Facts
Voice and AI: facts and numbers
322 facts, each traced to the exact words of its source.
- 9to5Google reported on September 28, 2026 that Google Assistant was gone on every device and account it checked in the US and Europe.Abner Li, 9to5Google, 2026
- 9to5Google reported on September 28, 2026 that users can no longer use or switch back to Google Assistant on their phone, tablet or paired devices.Abner Li, 9to5Google, 2026
- 9to5Google reported that the Switch to Google Assistant shortcut no longer appears at the bottom of the Gemini app's account menu in Google app 17.60 stable and 17.62 beta.Abner Li, 9to5Google, 2026
- A 2018 workshop paper on speech input in public spaces expected that users would be more self-conscious speaking their answers than typing them discreetly.Giovanna Nunes Vilaza et al., CHI 2018 Workshop on Socially Acceptable Technologies, UCL Interaction Centre, 2018
- A 2025 study of language models on domain-specific tasks found that longer prompts generally improved performance, while shorter prompts could make it worse.Qibang Liu et al., arXiv preprint 2502.14255, 2025
- A 2026 preprint reports that LLMs not only alter the voice and tone of human writing but also consistently alter its intended meaning.Marwa Abdulhai et al., arXiv, 2026
- 62 words per minuteA brain-computer interface decoded one participant's attempted speech into text at 62 words per minute, with a 23.8 percent word error rate on a 125,000-word vocabulary.Francis R. Willett et al., Nature, 2023
- 0.007 US dollars per hour of dictationA cleanup pass over one hour of dictated text, using Claude Haiku 5.5 prices, costs about 0.007 US dollars.Calculated from other facts
- 0.28 US dollars per hour of dictationA cleanup pass over one hour of dictated text, using Claude Opus 5.5 prices, costs about 0.28 US dollars.Calculated from other facts
- A January 2025 US executive order, Executive Order 14148, lists Executive Order 14110 of October 30, 2023, on safe, secure and trustworthy AI development, among the executive actions it revokes.Donald J. Trump, Compilation of Presidential Documents, Office of the Federal Register, via GovInfo, 2025
- A March 2026 9to5Google report says the redesigned voice input in the Gemini app does not submit a prompt automatically, which the article calls ideal for longer prompts.9to5Google, 2026
- 800 millisecondsA pipeline that runs Whisper large v3 first and then the Qwen3 8B language model had a 90th percentile time to first token of about 800 milliseconds on the same hardware.Dhruv Jain et al., arXiv, 2025
- 2,560 hours of audio per monthA rented GPU server that costs about 383.98 US dollars a month would have to transcribe about 2,560 hours of audio a month to cost the same as AssemblyAI's Universal-2 rate of 0.15 dollars per hour.Calculated from other facts
- A SamMobile article updated on August 6, 2026 reports that Google will start replacing Google Assistant with Gemini on Galaxy phones, tablets, watches, Buds and Android Auto on September 4, 2026.SamMobile, 2026
- 78 words per minuteA second brain-computer interface study decoded attempted silent speech into text at a median of 78 words per minute with a median word error rate of 25 percent.Sean L. Metzger et al., Nature, 2023
- 38.6 percent accuracy after one editA single adversarially chosen character edit lowered Mistral-7B-Instruct's accuracy on the GSM8K math benchmark from 43.7 percent to 38.6 percent.Esther Gan et al., arXiv preprint 2411.05345, 2024
- 15.12 US dollars per monthA user who dictates two hours on each of 21 workdays sends 2,520 minutes a month, which costs about 15.12 US dollars a month at gpt-4o-transcribe prices.Calculated from other facts
- According to Wispr Flow's privacy page, turning off context awareness means the app transcribes without reading anything on your screen.Wispr Flow (vendor website), 2026
- 208 millisecondsAcross ten languages, the mean gap between one speaker finishing a turn and the next speaker starting was 208 milliseconds.Tanya Stivers et al., Proceedings of the National Academy of Sciences, 2009
- 0.124 US dollars per minuteAdding Twilio's US outbound line rate to Retell's example total gives about 12.4 cents per minute for an agent that calls out on a real phone line.Calculated from other facts
- 19 percent word error rateAfter supervised fine-tuning, the word error rate on the erroneous sentences fell from 22 to 19 percent, and the authors report this across all the models they tested.Xiuwen Zheng et al., arXiv (ICASSP 2026), 2026
- Amazon describes Alexa+ as its personal AI assistant that gets things done.Charles Clark, Amazon, 2025
- Amazon introduces Alexa+ as the next generation of Alexa, powered by generative AI, and describes it as a personal AI assistant that gets things done.Charles Clark, Amazon, 2025
- Amazon researchers report that combining LLM prompting with fine-tuning brought speech recognition error rates below the N-best oracle level, a benchmark for choosing among the recognizer's candidate transcripts.Chao-Han Huck Yang et al., IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) 2023, 2023
- Amazon says Alexa+ can connect with products and services from Philips Hue, Uber Eats, OpenTable and Spotify.Charles Clark, Amazon, 2025
- Amazon says Alexa+ can extract key details from documents and add appointments to your calendar.Amazon (About Amazon), n.d.
- Amazon says Alexa+ can help you complete tasks such as managing your to-do list, updating your family calendar, controlling your smart home and making reservations.Amazon (About Amazon), n.d.
- $19.99 per monthAmazon says Alexa+ costs $19.99 per month and that all Amazon Prime members get it for free (page updated July 21, 2026).Charles Clark, Amazon, 2025
- Amazon says Alexa+ is available in a web browser at alexa.com.Amazon (About Amazon), n.d.
- Amazon Web Services defines an intent in Amazon Lex as something that represents an action that the user wants to perform.Amazon Web Services, Amazon Web Services, n.d.
- Amazon's help page for Remember This, the Alexa+ feature that stores notes you ask it to keep, says there is no limit to the number of notes you can create.Amazon Alexa Help, n.d.
- 13.5 percentAn excess vocabulary analysis of PubMed abstracts suggests that at least 13.5 percent of 2024 abstracts were processed with LLMs.Dmitry Kobak et al., arXiv, 2024
- 0.10 US dollars per million input tokensAnthropic lists Claude Haiku 5.5 at $0.10 per million input tokens for prompts up to 100,000 tokens.Anthropic, Anthropic, 2026
- 0.50 US dollars per million output tokensAnthropic lists Claude Haiku 5.5 at $0.50 per million output tokens.Anthropic, Anthropic, 2026
- 20 US dollars per million output tokensAnthropic lists Claude Opus 5.5 at $20 per million output tokens.Anthropic, Anthropic, 2026
- 4 US dollars per million input tokensAnthropic lists Claude Opus 5.5 at $4 per million input tokens.Anthropic, Anthropic, 2026
- Anthropic's official prompting guidance says to think of Claude as a capable new employee who lacks context, and that the more precisely you explain what you want, the better the result.Anthropic, Anthropic, 2026
- Apple and Google said in January 2026 that Gemini-based models will help power a more personalized Siri coming that year.Google and Apple, Google, 2026
- Apple says Apple Intelligence can work with ChatGPT from OpenAI, when you allow it, to help you get things done.Apple Support (Mac User Guide, macOS 27), 2026
- Apple says Apple Intelligence is not available on all Mac models or in all languages or regions.Apple, Apple Support, Mac User Guide (macOS Sequoia 15), 2026
- Apple says Apple Intelligence will keep running on Apple devices and Private Cloud Compute under the Google deal.Google and Apple, Google, 2026
- Apple says that with Siri AI, Siri gains personal context understanding, onscreen awareness and broad world knowledge.Apple Support, Mac User Guide, 2026
- Apple says you can also summarize an audio recording in Voice Memos on a Mac, using Writing Tools.Apple, Apple Support, Mac User Guide (macOS Sequoia 15), 2026
- Apple says you can record audio in the Notes app and then use Apple Intelligence to generate a summary of the transcript.Apple, Apple Support, Mac User Guide (macOS Sequoia 15), 2026
- 3 billion parametersApple's 2025 foundation models update describes a compact on-device model of approximately 3 billion parameters, optimized to run efficiently on Apple silicon.Apple Machine Learning Research, 2025
- Apple's Mac Help lists reading and writing text messages and emails among the things you can ask Siri to do on a Mac.Apple Support, Mac User Guide, 2026
- Apple's Mac Help refers to the new Siri capabilities as Siri AI (Beta).Apple Support, Mac User Guide, 2026
- Artificial Analysis defines time to first token for language model APIs as the number of seconds between sending a request and receiving the first token of the response. This is the delay a user waits before any words of a reply appear.Artificial Analysis, Artificial Analysis, 2026
- 9.6 percent of submissionsAs of 14 May 2023, Turnitin reported that 9.6% of the 38.5 million submissions it had processed for AI writing reported over 20% AI writing.Annie Chechitelli, Turnitin, 2023
- At Gemini's launch on Android in February 2024, Google listed alarms and timers, phone calls, and sending and reading messages among the Assistant actions Gemini could perform.Abner Li, 9to5Google, 2024
- At Gemini's launch on Android in February 2024, Google said podcasts, news and radio stations, and third-party music providers were not supported in Gemini.Abner Li, 9to5Google, 2024
- At Gemini's launch on Android in February 2024, Google said that to set reminders and tasks you had to switch back to Google Assistant.Abner Li, 9to5Google, 2024
- 0.24 US dollars per hour of audioAt Google's rate for 2,000,000 minutes and above each month, one hour of audio costs about 0.24 US dollars.Calculated from other facts
- 0.96 US dollars per hour of audioAt Google's standard rate for its first 500,000 minutes each month, one hour of audio costs about 0.96 US dollars.Calculated from other facts
- 70 percent of Home Assistant usersAt its launch, Speech-to-Phrase supported English, French, German, Dutch, Spanish and Italian, which Home Assistant said covers nearly 70 percent of its users.Home Assistant project, Home Assistant blog, 2025
- 0.27 US dollars per hour of audioAt OpenAI's listed price of $0.0045 per minute, one hour of audio transcribed with gpt-transcribe costs about $0.27.Calculated from other facts
- 1.02 US dollars per hour of audioAt OpenAI's listed price of $0.017 per minute for live transcription, one hour of audio with gpt-realtime-whisper costs about $1.02.Calculated from other facts
- 0.18 US dollars per hour of audioAt OpenAI's listed price, one hour of audio transcribed with gpt-4o-mini-transcribe costs about 0.18 US dollars.Calculated from other facts
- 0.36 US dollars per hour of audioAt OpenAI's listed price, one hour of audio transcribed with gpt-4o-transcribe costs about 0.36 US dollars.Calculated from other facts
- 0.082 to 0.129 US dollars per minuteAt Vapi's 1,000-minute example, the estimated total works out to about 8.2 to 12.9 cents per minute.Calculated from other facts
- Athenaeum's AI policy does not allow students to have an AI tool rephrase entire sentences or paragraphs of their writing.Athenaeum, n.d.
- 7.08 US dollars per monthBilled once a year, Superwhisper's Pro plan works out to about 7.08 US dollars a month.Calculated from other facts
- Birchtree reports that Advanced Dictation, Apple's AI dictation feature in the iOS 27 beta, is off by default.Birchtree, 2026
- Burren College's generative AI policy says that for the tools it excludes from its definition of AI, students must still be able to declare unequivocal authorship of their work.Burren College of Art, 2026
- 0.0077 US dollars per minuteDeepgram lists its Nova-3 monolingual model at $0.0048 per minute at the current price and $0.0077 per minute at the regular price.Deepgram, 2026
- Dictionary.com gives a verb sense of dictate: to prescribe or lay down authoritatively or peremptorily, which it glosses as to command unconditionally.Dictionary.com, Dictionary.com, n.d.
- Dictionary.com lists dictates, in the plural, as a noun meaning a guiding or governing principle or requirement, and gives the example of following the dictates of one's conscience.Dictionary.com, Dictionary.com, n.d.
- Dictionary.com says dictate can also be used in a broader sense, and gives weather as an example: weather often dictates how many layers of clothes you wear.Dictionary.com, Dictionary.com, n.d.
- Dictionary.com traces dictate to the Latin verb dictāre, which it glosses as to say repeatedly, order, prescribe.Dictionary.com, Dictionary.com, n.d.
- 2.7 timesEnglish speech input on a phone, at 153 words per minute, is about 2.7 times the 2026 home brain-computer interface average of 56 words per minute.Calculated from other facts
- Even when an LLM is asked only for grammar edits, the same 2026 preprint finds, it still changes the text in a way that significantly alters its semantic meaning.Marwa Abdulhai et al., arXiv, 2026
- Feeding a language model word confusion networks from the speech recognizer's lattice, instead of only its top transcript, helped bridge the gap to an oracle upper bound in spoken question answering and intent classification.Kevin Everson et al., arXiv preprint 2401.02921, 2024
- For smart-home actions that Gemini does not support, the Gemini app opens the Google Home app or gives you a link to it so you can control the device there.Google, n.d.
- 9.14 US dollars per hour of audioFor the illustrative heavy user who dictates 42 hours a month, an always-on GPU server costs about 9.14 US dollars for each hour of dictated audio.Calculated from other facts
- 17.7 monthsFrom Google's March 14, 2025 announcement to the September 4, 2026 start date reported for Gemini replacing Assistant is about 17.7 months.Calculated from other facts
- Google Home works with Matter, a smart home standard that helps devices from different brands work together.Google, n.d.
- Google said in its blog post that it was upgrading more mobile users from Google Assistant to Gemini over the coming months.Google Blog, 2025
- 6 languagesGoogle says Assistant voice typing is available in six languages: English, French, German, Italian, Japanese and Spanish.Google, n.d.
- Google says basic Gemini for Home features are free, including smart home controls, alarms and timers, calendars, notes, lists and reminders.Google, n.d.
- Google says Gemini now answers actions such as weather with voice and text only, and no longer shows visual information cards.Google, n.d.
- Google says Gemini will replace Google Assistant on your paired mobile devices, including smart watches, headphones and earbuds with Google Assistant built in.Google, n.d.
- Google says Google Assistant keeps responding to Hey Google on non-mobile devices, including smart displays, smart speakers, TVs and cars.Google, n.d.
- Google says some familiar Google Assistant features may not be supported in Gemini, which is still evolving.Google, n.d.
- Google says that Gemini Live on compatible devices needs a Google Home Premium subscription, while basic smart-home controls do not.Google, n.d.
- Google says that home controls in Gemini are for convenience only, not for safety-critical or security-critical purposes.Google, n.d.
- Google says that if you are signed in to a Google Workspace for Education account on an Android device and cannot turn on Hey Google, your admin may have turned off settings that Hey Google needs.Google, 2026
- Google says that on Android, the Gemini app gets help from Google Assistant for some familiar quick voice action features.Google, n.d.
- Google says that once a home switches to Gemini for Home voice assistant, it cannot switch back to Google Assistant.Google, n.d.
- Google says that once you set up a Google smart speaker or display in the Google Home app, you can use voice commands to control your linked smart devices.Google, n.d.
- Google says that the text you speak with Assistant voice typing stays on your device and is not sent to Google servers.Google, n.d.
- 14 languagesGoogle says the Hey Google and quick voice action features with Gemini are not available in 14 languages, including Greek, Hebrew and Ukrainian.Google, n.d.
- 50,000 smart home devicesGoogle says you can control over 50,000 smart home devices, such as TVs, lights, appliances, plugs and thermostats, once you add them to the Google Home app.Google, n.d.
- Google's examples for a phone include calling a contact, and answering or rejecting an incoming call by voice.Google, n.d.
- Google's help page says some starters and actions are not supported when you use a routine with Gemini, and Gemini tells you when an action is not supported.Google, n.d.
- Google's help page says that when you switch to Gemini on Android, Gemini replaces Google Assistant as the primary assistant on your device.Google, n.d.
- Google's help page says you can switch your phone or tablet back to Google Assistant as the primary assistant in the Gemini app settings.Google, n.d.
- Google's list of examples for a speaker or Smart Display includes smart-home commands for lights and thermostats, such as dimming the living room lights and setting the heat to 70.Google, n.d.
- Google's Machine Learning Glossary defines a prompt as any text entered as input to a large language model to condition the model to behave in a certain way.Google Developers, Google Developers, 2026
- Google's Machine Learning Glossary defines machine learning as a program or system that trains a model from input data, and the trained model can then make predictions from new data.Google Developers, Google Developers, 2026
- Google's Machine Learning Glossary describes generative AI as an emerging field with no formal definition.Google Developers, Google Developers, 2026
- Google's Machine Learning Glossary says that, at a minimum, a large language model is a language model having a very high number of parameters.Google Developers, Google Developers, 2026
- Google's March 2025 blog post said that later in 2025, the classic Google Assistant would no longer be accessible on most mobile devices.Google Blog, 2025
- 20 US dollars per monthGoogle's October 2025 announcement priced the Google Home Premium Advanced plan at $20 a month, or $200 a year.Google, The Keyword blog, 2025
- 10 US dollars per monthGoogle's October 2025 announcement priced the Google Home Premium Standard plan at $10 a month, or $100 a year.Google, The Keyword blog, 2025
- Google's steps for Hey Google on an Android phone: open the Google Assistant app, say Assistant settings, then under Popular settings tap Hey Google and Voice Match.Google, n.d.
- Google's troubleshooting checklist for smart home devices asks whether your devices are connected to the same Wi-Fi network.Google, n.d.
- Home Assistant describes Piper as a fast, local neural text-to-speech system that is optimized for the Raspberry Pi 4.Home Assistant project, Home Assistant, 2026
- Home Assistant describes Speech-to-Phrase as a close-ended speech model that transcribes only what it knows, with support for only a subset of Assist's voice commands.Home Assistant project, Home Assistant, 2026
- Home Assistant describes Whisper as an open-ended speech model that tries to transcribe everything, at the cost of slower processing.Home Assistant project, Home Assistant, 2026
- Home Assistant lists an Alexa Smart Home Skill as one way to use Alexa with Home Assistant, to control lights and more.Home Assistant, 2026
- Home Assistant recommends at least an Intel N100 or an equivalent processor for running OpenAI's Whisper locally, and says the Base model runs reasonably fast for English and Spanish.Home Assistant project, Home Assistant, 2026
- 25 entitiesHome Assistant recommends exposing fewer than 25 entities to a local language model when you experiment with Ollama, and warns that smaller models make more mistakes.Home Assistant project, Home Assistant, 2026
- 150 milliseconds per commandHome Assistant reports that the Raspberry Pi 5 processes Speech-to-Phrase commands seven times faster than the Raspberry Pi 4, at 150 milliseconds per command.Home Assistant project, Home Assistant blog, 2025
- Home Assistant says a standardized hardware platform for voice is an essential ingredient for the community to drive its voice project forward, and that the Voice Preview Edition was built for that purpose.Home Assistant project, Home Assistant, 2026
- Home Assistant says Assist handles commands first by default, and only sends questions or commands it does not understand to the AI you have set up.Home Assistant project, Home Assistant blog, 2025
- Home Assistant says delayed commands, such as turning off the lights in 5 minutes, cannot be cancelled or modified, unlike regular voice timers.Home Assistant project, Home Assistant, 2026
- 5 voice satellitesHome Assistant says five voice satellites can stream audio at the same time without overwhelming a Raspberry Pi 4, because wake word detection runs inside Home Assistant.Home Assistant project, Home Assistant, 2026
- Home Assistant says French may work well enough with the largest Whisper model, which needs a GPU, but is unusable on a Raspberry Pi or even an N100-class PC.Home Assistant project, Home Assistant blog, 2025
- Home Assistant says Home Assistant Cloud is an optional subscription that supplies the speech engines used by Assist, and that Home Assistant works fully without it.Home Assistant project, Home Assistant, 2026
- Home Assistant says its cloud speech processing runs on Microsoft Azure under commercial terms, and that Home Assistant Cloud itself keeps no record of your voice, data or commands.Home Assistant project, Home Assistant, 2026
- Home Assistant says its Cloud subscription is planned to be renamed Home Assistant Link on December 2, 2026.Home Assistant project, Home Assistant, 2026
- Home Assistant says only Ollama models that support tools may control Home Assistant, so a model without tool support can answer questions but cannot act on devices.Home Assistant project, Home Assistant, 2026
- 1 secondHome Assistant says Speech-to-Phrase, a closed-ended model that transcribes only what it knows, transcribes in under one second even on a Home Assistant Green or Raspberry Pi 4.Home Assistant project, Home Assistant, 2026
- Home Assistant says that before it added the prefer-local setting, many easy-to-run commands were being sent to a language model, which slowed responses and wasted tokens.Home Assistant project, Home Assistant blog, 2025
- Home Assistant says that for some languages it believes Home Assistant Cloud provides the best experience, because the local speech-to-text models do not support them.Home Assistant project, Home Assistant, 2026
- Home Assistant says that Home Assistant Cloud, the route to automatic Google Assistant setup, requires a paid subscription after a 30-day free trial.Home Assistant, 2026
- Home Assistant says that manual setup of the Amazon Alexa integration has several requirements, including an Amazon Developer Account.Home Assistant, 2026
- 1.6 seconds of speech per secondHome Assistant says that on a Raspberry Pi, using medium quality Piper models, the text-to-speech engine can generate 1.6 seconds of voice in one second.Home Assistant project, Home Assistant, 2026
- Home Assistant says that once the speech-to-text and text-to-speech apps are running, the Wyoming integration discovers both services, and you select Add for each.Home Assistant project, Home Assistant, 2026
- Home Assistant says that shopping lists, naming a timer and broadcasts are not usable with Speech-to-Phrase out of the box.Home Assistant project, Home Assistant, 2026
- Home Assistant says that the speech-to-text and text-to-speech options that run entirely local send no data to external servers for processing.Home Assistant project, Home Assistant, 2026
- Home Assistant says that to run the whole local voice setup together, you need the Home Assistant Operating System running.Home Assistant project, Home Assistant, 2026
- Home Assistant says that until you expose your devices to Assist, you won't be able to control them by voice.Home Assistant project, Home Assistant, 2026
- Home Assistant says that, to use Google Assistant without Home Assistant Cloud, your configuration has to be reachable from outside with a hostname and an SSL certificate.Home Assistant, 2026
- Home Assistant says that, with Home Assistant Cloud, you can connect your Home Assistant instance to Amazon Alexa in a few simple clicks.Home Assistant, 2026
- Home Assistant says the mute switch on the Voice Preview Edition physically cuts power to the microphones.Home Assistant project, Home Assistant, 2026
- Home Assistant says the OpenAI integration requires an API key and is a paid service, so you should monitor your costs in the OpenAI portal.Home Assistant project, Home Assistant, 2026
- Home Assistant says timers created with Assist are voice timers that belong to the device where you started them, and that they do not create timer entities you can use in automations.Home Assistant project, Home Assistant, 2026
- Home Assistant says Whisper is only a great option if you have powerful hardware at home and plan to extend your voice setup beyond simple home control, for example by pairing it with an LLM-based agent.Home Assistant project, Home Assistant, 2026
- Home Assistant says you can add your own sentences to Assist's intent recognizer, either by extending an existing intent or by creating a new one.Home Assistant project, Home Assistant, 2026
- Home Assistant says you cannot use a language model as a fallback with Speech-to-Phrase, because unexpected commands cannot be converted into text for the model to process.Home Assistant project, Home Assistant blog, 2025
- 8k tokens of context by defaultHome Assistant's default context window for the Ollama integration is 8k tokens, which is larger than the default of 2k on the Ollama server.Home Assistant project, Home Assistant, 2026
- Home Assistant's documentation says the OpenAI conversation integration does not integrate with sentence triggers.Home Assistant project, Home Assistant, 2026
- 8 secondsHome Assistant's documentation says Whisper takes around 8 seconds to process a voice command on a Raspberry Pi 4 and under a second on an Intel NUC.Home Assistant project, Home Assistant, 2026
- 50 percentHome Assistant's opt-in analytics show that over 50 percent of Home Assistant OS users run their homes on affordable, low-powered machines such as the Home Assistant Green or a Raspberry Pi.Home Assistant project, Home Assistant blog, 2025
- 69 US dollarsHome Assistant's product page shows the Voice Preview Edition at $69 in US dollars, or 59 euros.Home Assistant project, Home Assistant, 2026
- December 2024 month of launchHome Assistant's timeline of its voice work lists the launch of the Voice Preview Edition in December 2024.Home Assistant project, Home Assistant, 2026
- In a 2020 human evaluation, annotators found substantial amounts of hallucinated content in every summary that the neural abstractive models generated.Joshua Maynez et al., Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020
- In a 2023 evaluation of ChatGPT as a grammar corrector, its very fluent corrections were possibly due to over-correction and not following the principle of minimal edits.Tao Fang et al., arXiv, 2023
- In a 2023 study of several widely used GPT detectors, the detectors consistently misclassified non-native English writing samples as AI-generated.Weixin Liang et al., arXiv (later published in Patterns), 2023
- 31.6 words per minuteIn a 2024 BrainGate study, one participant's copy task was decoded at his self-paced speaking rate of 31.6 words per minute.Nicholas S. Card et al., medRxiv preprint, deposited in PubMed Central (later published in the New England Journal of Medicine, 2024), 2024
- In a 2024 declaratory ruling, the FCC said the TCPA limits on artificial or prerecorded voices cover current AI technologies that generate human voices.Federal Communications Commission, 2024
- In a 2024 study, fine-tuned GPT-2 and T5 models were not robust to the errors of a speech recognizer by default, but training them on spoken dialogue data made them more robust.Seyed Mahed Mousavi et al., Proceedings of IWSDS 2024 (arXiv preprint 2401.02297), 2024
- 32 percent character error rateIn a 2025 preprint, the Brain2Qwerty decoder reached an average character error rate of 32 percent with magnetoencephalography (MEG) recordings.Jarod Levy et al., arXiv (preprint), 2025
- In a 2025 technical report, Chroma found that across all its experiments, model performance consistently degrades as the input gets longer, even on simple tasks.Kelly Hong et al., Chroma, 2025
- 49.7 words per minuteIn a 2026 copy task, the participant's silent speech was decoded at an average of 49.7 words per minute, with accuracy of 96.5 percent.Nicholas S. Card et al., Nature Medicine, 2026
- In a 2026 ICML study, many large language models kept notable performance even when most words were shuffled letter by letter or many invisible characters were inserted.Anyuan Zhuo et al., ICML 2026, 2026
- In a 2026 preprint on conversational speech, reasoning models systematically over-deleted fluent content, which the authors read as a bias toward abstraction over fidelity to the transcript.Maria Teleki et al., arXiv, 2026
- In a 2026 study of spoken questions about code, having a language model refine the speech recognizer's output significantly improved performance in both the transcription and the code understanding stages.Jayant Havare et al., arXiv, 2026
- 44 percent average transcription error rateIn a 2026 study, 15 speech models from OpenAI, Deepgram, Google and Microsoft had an average transcription error rate of 44 percent on US street names spoken by US participants.Kaitlyn Zhou et al., arXiv, 2026
- In a car, Google's examples for communication include calling Mom, texting Dad and reading your messages aloud.Google, n.d.
- In a controlled experiment on argumentative essays, writing with the InstructGPT model produced a statistically significant reduction in diversity, while writing with the base GPT3 model did not.Vishakh Padmakumar and He He, International Conference on Learning Representations (ICLR) 2024; arXiv, 2024
- 35 percentIn a February 2026 Pew Research Center survey of U.S. adults, 35 percent said they have a smart speaker, ahead of a smart doorbell (18 percent) and a robot vacuum (13 percent).Pew Research Center, Pew Research Center, 2026
- In a joint statement dated January 12, 2026, Apple and Google said the next generation of Apple Foundation Models will be based on Google's Gemini models and cloud technology, helping power a more personalized Siri.Google and Apple, Google, 2026
- In a March 2025 post, Google said the classic Google Assistant would stop being accessible on most mobile devices later that year.Google Blog, 2025
- 51.56 words per minuteIn a study of 136 million keystrokes from an online typing test, the average typing speed of participants was 51.56 words per minute.Vivek Dhakal et al., CHI 2018, ACM Conference on Human Factors in Computing Systems, 2018
- 3 percentIn a survey reported in June 2016, 3 percent of iPhone users said they had used Siri in public.GSMArena, 2016
- In April 2026, MacRumors reported that it was still unclear whether the new Siri and Gemini-powered Apple Intelligence features would run on Apple's Private Cloud Compute or on Google's servers.MacRumors, 2026
- In August 2024, Gemini on Android began running Google Assistant Routines when you say start and the routine name, but typed triggers did not work.Abner Li, 9to5Google, 2024
- In August 2024, Google said it used Gemini Nano with multimodality to further improve the Recorder app's summarization feature on Pixel 9 devices.Terence Zhang and Kristi Bradford, Android Developers Blog (Google), 2024
- In August 2024, Google said that Pixel Recorder's audio summaries use Gemini Nano and keep data processing on the device.Terence Zhang and Kristi Bradford, Android Developers Blog (Google), 2024
- In December 2023, 9to5Google reported that Pixel Recorder's Summarize feature ran on the Pixel 8 Pro using Gemini Nano, while other Pixel devices received Transcribe again.Abner Li, 9to5Google, 2023
- In February 2024 the Federal Communications Commission ruled unanimously that voices generated by artificial intelligence are artificial voices under the Telephone Consumer Protection Act.Cooley LLP, Cooley LLP, 2024
- 5 secondsIn February 2025, Home Assistant said that Whisper on a Raspberry Pi 4 takes at least 5 seconds to turn speech into text, which is why it recommended at least an Intel N100 for fully local voice.Home Assistant project, Home Assistant blog, 2025
- 76 accuracy pointsIn few-shot prompts that differed only in formatting, LLaMA-2-13B's accuracy varied by up to 76 percentage points, according to the study.Melanie Sclar et al., ICLR 2024 (arXiv preprint 2310.11324), 2024
- In Home Assistant's interface, you create a sentence trigger by entering one or more sentences, and you leave out punctuation.Home Assistant project, Home Assistant, 2026
- In logs of a 2010 mobile search service, longer queries had a higher probability of being typed than shorter ones, which the authors said was contrary to their expectation.Maryam Kamvar and Doug Beeferman, Interspeech 2010, ISCA, 2010
- In March 2025 Google announced that it would replace Google Assistant on Android phones with Gemini, and that Assistant would no longer be accessible on most mobile devices or available from app stores later that year.Aisha Malik, TechCrunch, 2025
- 6,000,000 US dollarsIn September 2024 the FCC imposed a $6,000,000 penalty on Steve Kramer for an AI-voice robocall campaign that targeted New Hampshire voters.Federal Communications Commission, 2024
- In tests on long inputs, language model accuracy was highest when the relevant information sat at the start or end of the input, and dropped significantly when it sat in the middle.Nelson F. Liu et al., arXiv (accepted to Transactions of the Association for Computational Linguistics), 2023
- 92 percentIn the 2018 AlterEgo paper, a wearable that reads silently mouthed words reached a median word accuracy of 92 percent in its user studies.Arnav Kapur et al., 23rd International Conference on Intelligent User Interfaces (IUI 2018), 2018
- 6.3 correct words per minuteIn the 2024 BrainGate study, the participant typed with a gyroscopic head mouse at 6.3 correct words per minute, with a standard deviation of 1.3.Nicholas S. Card et al., medRxiv preprint, deposited in PubMed Central (later published in the New England Journal of Medicine, 2024), 2024
- 97.5 percent accuracyIn the 2024 BrainGate study, the speech neuroprosthesis sustained 97.5 percent accuracy beyond eight months after surgery.Nicholas S. Card et al., medRxiv preprint, deposited in PubMed Central (later published in the New England Journal of Medicine, 2024), 2024
- In the 2025 INTERSPEECH paper, existing language-model error correction for speech recognition relies mainly on the text, ignores phonetic cues, and that leads to over-correction, meaning the model changes words that were already right.Natsuo Yamashita et al., arXiv (accepted by INTERSPEECH 2025), 2025
- 10 millisecondsIn the 2025 Nature study, the decoder turned the participant's neural activity into voice causally, within 10 milliseconds.Maitreyee Wairagkar et al., Nature, 2025
- In the 2025 Nature study, the voice synthesized from the participant's neural activity was often intelligible, but not consistently so.Maitreyee Wairagkar et al., Nature, 2025
- 62 percentIn the 2026 copy task, silent speech at 49.7 words per minute was about 62 percent faster than vocalized speech at 30.6 words per minute.Calculated from other facts
- 56 words per minuteIn the 2026 Nature Medicine study, the participant communicated with the speech BCI at home at an average rate of 56 words per minute over almost two years of use.Nicholas S. Card et al., Nature Medicine, 2026
- 92 percent of sentencesIn the 2026 Nature Medicine study, the participant labeled 92 percent of his sentences as decoded at least mostly correctly.Nicholas S. Card et al., Nature Medicine, 2026
- 3,800 hoursIn the 2026 Nature Medicine study, the participant used the brain-computer interface at home for more than 3,800 hours with no researchers present.Nicholas S. Card et al., Nature Medicine, 2026
- In the Home Assistant OpenAI integration, storing requests and responses in the OpenAI dashboard is off by default.Home Assistant project, Home Assistant, 2026
- 77.1 percent relative reduction in word error rateIn the HyPoradise benchmark, the H2T-LoRA language model correction reduced word error rate by 77.1 percent relative on the ATIS test set and by 55.1 percent relative on WSJ.Chen Chen et al., arXiv (NeurIPS 2023 Datasets and Benchmarks track), 2023
- 1.3 percentIn the same 2016 survey report, 1.3 percent of iPhone users said they had used Siri at work.GSMArena, 2016
- In the same 2024 experiment, access to generative AI ideas made stories rated as more creative, better written and more enjoyable, especially among less creative writers.Anil R. Doshi and Oliver P. Hauser, Science Advances, 2024
- In the same 2024 experiment, stories written with generative AI ideas were more similar to one another than stories written by humans alone.Anil R. Doshi and Oliver P. Hauser, Science Advances, 2024
- 30.6 words per minuteIn the same 2026 copy task, the participant's vocalized speech was decoded at an average of 30.6 words per minute, with accuracy above 99 percent.Nicholas S. Card et al., Nature Medicine, 2026
- 0.45 word error rate for Black speakersIn the study, Apple's system had a word error rate of 0.45 for Black speakers and 0.23 for white speakers.Allison Koenecke et al., Proceedings of the National Academy of Sciences, 2020
- 0.27 word error rate for Black speakersIn the study, Microsoft's system had a word error rate of 0.27 for Black speakers and 0.15 for white speakers.Allison Koenecke et al., Proceedings of the National Academy of Sciences, 2020
- 20 timesIn the VoiceAgentBench measurements, the two-stage pipeline's 90th percentile time to first token was about 20 times that of the end-to-end speech model.Calculated from other facts
- In Word's dictation settings, the Filter sensitive phrases option masks potentially sensitive phrases with asterisks, and it can be switched off.Microsoft, 2026
- 1.0 secondJakob Nielsen's response-time limits say 1.0 second is about the limit for the user's flow of thought to stay uninterrupted, although the user still notices the delay.Jakob Nielsen, Nielsen Norman Group, 1993
- 1972 yearJurafsky and Martin's 2026 draft says hidden Markov models seem to have been applied to speech independently at two laboratories around 1972.Daniel Jurafsky and James H. Martin, Stanford University, online manuscript of the third edition, 2026
- Microsoft describes its speech to text service as offering versatile solutions for converting audio streams into text, through both real-time and batch transcription.Microsoft, Microsoft Learn, n.d.
- Microsoft says fluid dictation in Windows Voice Access automatically corrects grammar, punctuation and filler words as you speak.Microsoft Support, 2026
- Microsoft says the fluid dictation experience in Windows Voice Access is powered by on-device small language models (SLMs).Microsoft Support, 2026
- Microsoft says Windows voice typing uses online speech recognition, and that the online recognition is powered by Azure Speech services.Microsoft, 2026
- Mistral's Voxtral models are trained to understand both spoken audio and text documents, and the paper calls them multimodal audio chat models.Mistral AI, arXiv (Audio and Speech Processing), 2025
- 200 millisecondsMoshi, a full-duplex speech-to-speech model from Kyutai, has a theoretical latency of 160 ms and about 200 ms in practice.Alexandre Défossez and and colleagues at Kyutai, arXiv (Kyutai), 2024
- NIST's 2024 Generative AI Profile defines generative AI as the class of AI models that emulate the structure and characteristics of input data in order to generate derived synthetic content, which it says can include audio and text.National Institute of Standards and Technology, National Institute of Standards and Technology, U.S. Department of Commerce, 2024
- NVIDIA's Canary-Qwen model maps the audio encoder's output into the language model's embedding space through a linear projection.NVIDIA, Hugging Face, n.d.
- Ollama's FAQ says that when you use its cloud-hosted models, it processes your prompts and responses to provide the service, and does not store or log that content.Ollama, 2026
- 13.63 percent word error rateOn a dysarthric speech test set, the recognizer alone had an overall word error rate of 13.63 percent: 0 percent on sentences it transcribed without error, and 21.98 percent on sentences that had errors.Xiuwen Zheng et al., arXiv (ICASSP 2026), 2026
- On a Galaxy phone, a summary of a recording is opened from the Transcript assist icon, then the Summarize option.Samsung Support (US), n.d.
- more than 8 times slowerOn a Raspberry Pi 4, Whisper takes about 8 seconds per command and Speech-to-Phrase takes under one second, so Whisper is more than eight times slower.Calculated from other facts
- 64 percent higher word error rateOn FLEURS Arabic, the word error rate of Gemini 2.5 Flash was about 64 percent higher than that of Whisper large-v3, in the Voxtral paper's comparison.Calculated from other facts
- 57 percent lower word error rateOn FLEURS Hindi, the word error rate of GPT-4o mini Transcribe was about 57 percent lower than that of Whisper large-v3, in the Voxtral paper's comparison.Calculated from other facts
- 26 percentOn OpenAI's challenge set of English texts, its AI text classifier correctly identified 26 percent of AI-written text as likely AI-written.Jan Hendrik Kirchner et al., OpenAI, 2023
- 9 percentOn OpenAI's challenge set, its AI text classifier incorrectly labeled 9 percent of human-written text as AI-written.Jan Hendrik Kirchner et al., OpenAI, 2023
- 32.5 percentOn the 2012 Switchboard task, the move from a GMM-HMM baseline at 27.4 percent word error rate to a deep network at 18.5 percent is a relative reduction of about 32.5 percent.Calculated from other facts
- 24 percent lower word error rateOn the Open ASR Leaderboard's English short-form tests, the average word error rate of Canary-Qwen 2.5B was about 24 percent lower than that of Whisper Large v3.Calculated from other facts
- 24.2 percentOn the same Vectara leaderboard table, the highest hallucination rate is 24.2 percent, for the mistralai/ministral-3-3b-2512 model, which answered only 74.3 percent of prompts.Vectara, Vectara, GitHub, 2026
- 1.8 percentOn the Vectara leaderboard updated September 22, 2026, the lowest hallucination rate in the table is 1.8 percent, for the antgroup/finix_s1_32b model.Vectara, Vectara, GitHub, 2026
- 85 percent task completionOn the τ-Voice benchmark, GPT-5 with reasoning completed 85 percent of tasks, against 31 to 51 percent for voice agents on clean audio.Soham Ray et al., arXiv, 2026
- 26 to 38 percent task completionOn the τ-Voice benchmark, voice agents completed 26 to 38 percent of tasks under realistic conditions, with noise, diverse accents and interruptions.Soham Ray et al., arXiv, 2026
- Once you connect the Google Home app to the Gemini mobile app, Gemini can control the smart home devices on your Google Account, including devices you add later.Google, n.d.
- OpenAI Academy says that in the ChatGPT mobile or desktop app, you start Voice Mode by tapping and holding the large microphone button, which begins a real-time conversation, and you release it to stop.OpenAI, OpenAI Academy, 2026
- OpenAI describes its speech-to-speech path as a single model that takes audio in, decides what to do and speaks back.OpenAI, 2026
- 64.00 US dollars per 1M audio output tokensOpenAI lists gpt-realtime-2.1 audio output at $64.00 per 1M tokens and audio input at $32.00 per 1M tokens.OpenAI, OpenAI, 2026
- 0.05 US dollars per minuteOpenAI lists its gpt-live-1 voice sessions at $0.05 per minute, billed per second, with the backend model and tool usage charged separately.OpenAI, OpenAI, 2026
- 1,000 charactersOpenAI said its AI text classifier was very unreliable on short texts, below 1,000 characters.Jan Hendrik Kirchner et al., OpenAI, 2023
- OpenAI says ChatGPT Voice transcripts are not verbatim records and may not exactly match what was said.OpenAI, OpenAI Help Center, 2026
- OpenAI says the chained path is the one to use when you want to inspect or change the text between speech recognition, the agent and speech generation.OpenAI, 2026
- OpenAI says to use ChatGPT Voice for a live, back-and-forth conversation or to talk through ideas.OpenAI, OpenAI Help Center, 2026
- OpenAI withdrew its AI text classifier on July 20, 2023, because of its low rate of accuracy.Jan Hendrik Kirchner et al., OpenAI, 2023
- OpenAI's README for Whisper says that its code and model weights are released under the MIT License, a permissive license that lets anyone use them, including for free.OpenAI, GitHub (openai/whisper), 2026
- OpenWhispr says its cleanup fixes obvious grammatical slips and otherwise leaves the wording alone.OpenWhispr, OpenWhispr, n.d.
- OpenWhispr's own documentation says its default cleanup means that, occasionally, the text is not quite what you said.OpenWhispr, OpenWhispr, n.d.
- 3 built-in wake wordsOut of the box, the Voice Preview Edition listens for three wake words, Okay Nabu, Hey Jarvis and Hey Mycroft, using the on-device engine microWakeWord.Home Assistant project, Home Assistant, 2026
- Pu and colleagues say that when an LLM corrects speech recognition output, it makes many unnecessary changes to the input transcription, which they call over-correction.Jie Pu et al., arXiv, 2023
- 40 millisecondsQwen2.5-Omni 7B, an end-to-end speech language model, had a 90th percentile time to first token of about 40 milliseconds on one H100 GPU.Dhruv Jain et al., arXiv, 2025
- 0.11 US dollars per minuteRetell's pricing calculator example totals $0.11 per minute, made up of an LLM cost, voice infrastructure and text-to-speech, with telephony at $0.00.Retell AI, 2026
- 383.98 US dollars per monthRunning the g4dn.xlarge GPU instance all month, around the clock, costs about 383.98 US dollars at its on-demand rate.Calculated from other facts
- SamMobile reports that once Google Assistant is removed from a device that supports Gemini, the user cannot access it or switch back to it.SamMobile, 2026
- Samsung's Voice Recorder with Galaxy AI suggests actions from a transcribed recording, such as adding contacts, events or reminders for tasks.Samsung Support (US), n.d.
- 35 percentSmart speaker ownership in the United States rose only from 33 percent in 2021 to 35 percent in 2025, which the publisher describes as a plateau.Edison Research, Edison Research (now SSRS), 2025
- 3.0 timesSpeech input at 153 words per minute is about three times the 51.56 words per minute average from the online typing study. The two studies used different people and settings, so the ratio is a rough comparison.Calculated from other facts
- Spokenly says its optional on-device cleanup model removes filler words and repeats, applies corrections the speaker says out loud, fixes punctuation and capitals, and writes numbers as digits.Spokenly, Spokenly, n.d.
- 100 millisecondsStivers and colleagues report that speakers become hypersensitive to changes in the timing of responses of 100 milliseconds or less.Tanya Stivers et al., Proceedings of the National Academy of Sciences, 2009
- Superwhisper's documentation says its Voice to Text mode is transcription only, with no AI processing, so results arrive quickly.Superwhisper Documentation, 2026
- 3 context typesSuperwhisper's documentation says Super Mode is the only built-in mode with all three context types on by default.Superwhisper Documentation, 2026
- Superwhisper's documentation says that context awareness lets it process your dictation with relevant information from your screen.Superwhisper Documentation, 2026
- 2 modelsSuperwhisper's models page says it runs two models in sequence: a speech recognition model turns your voice into text, then an optional language model rewrites that text in Super Mode.Superwhisper, 2026
- Superwhisper's models page says its on-device transcription runs on the device, and that audio never leaves the device.Superwhisper, 2026
- Superwhisper's models page says that an on-device Parakeet speech model followed by a cloud Claude Sonnet language model is a common setup.Superwhisper, 2026
- Superwhisper's models page says that with cloud transcription, audio goes to the Superwhisper proxy, is transcribed and comes back.Superwhisper, 2026
- 29.4 monthsSuperwhisper's one-time lifetime price of 249.99 dollars equals about 29.4 months of its monthly price of 8.49 dollars.Calculated from other facts
- Superwhisper's Super Mode documentation says that application context reads text from the current input field or text editor.Superwhisper Documentation, 2026
- Superwhisper's Super Mode documentation says that when application context is active, the app also includes the current date and time, your full name and your computer name.Superwhisper Documentation, 2026
- Superwhisper's Super Mode documentation says you can open a dictation in History to review everything that was sent to the AI.Superwhisper Documentation, 2026
- The 1955 proposal for the Dartmouth summer project on artificial intelligence said an attempt would be made to find how to make machines use language.John McCarthy et al., Stanford University, John McCarthy's historical archive, 1955
- 175 billion parametersThe 2020 GPT-3 paper describes GPT-3 as an autoregressive language model with 175 billion parameters, 10 times more than any previous non-sparse language model.Tom B. Brown et al., Advances in Neural Information Processing Systems 33 (NeurIPS 2020); arXiv preprint, 2020
- The 2020 study found that error rates for Black speakers were nearly twice as large as for white speakers in every case.Allison Koenecke et al., Proceedings of the National Academy of Sciences, 2020
- The 2023 study describes an endpointer as a model that detects whether the user has finished speaking.Colin Lea et al., Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (arXiv preprint), 2023
- The 2023 study says that in dictation, the user finishes by clicking a button again, instead of relying on an endpointer.Colin Lea et al., Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (arXiv preprint), 2023
- The 2024 paper says end-to-end speech recognition systems still have a significant error rate for entity names that appear rarely in their training data.Ernest Pusateri and et al. (nine authors in total), arXiv, 2024
- The 2025 INTERSPEECH authors add the speech recognizer's N-best hypotheses (its ranked alternative transcripts) and phonetic context to the language model, to reduce over-correction. They report lower word and character error rates on English and Japanese datasets.Natsuo Yamashita et al., arXiv (accepted by INTERSPEECH 2025), 2025
- 9 percentThe 2026 home average of 56 words per minute is about 9 percent above the 51.56 words per minute average typist in a large online test.Calculated from other facts
- The 2026 Nature Medicine study of a speech and cursor brain-computer interface reports results from one participant, a man with paralysis and severe dysarthria due to ALS.Nicholas S. Card et al., Nature Medicine, 2026
- 77.6 yearsThe 680,000 hours of training audio used for Whisper add up to about 77.6 years of continuous listening.Calculated from other facts
- The authors found that the loss of diversity came mainly from the InstructGPT model's contributions to co-written essays, while the text the participants wrote themselves was not affected.Vishakh Padmakumar and He He, International Conference on Learning Representations (ICLR) 2024; arXiv, 2024
- 0.526 US dollars per hourThe AWS g4dn.xlarge instance, which has one NVIDIA T4 GPU, starts at $0.526 per hour on demand in the US East (N. Virginia) region, according to a price tracker updated October 16, 2026.Holori, 2026
- 35 healthy volunteersThe Brain2Qwerty preprint tests its decoder in a group of 35 healthy volunteers.Jarod Levy et al., arXiv (preprint), 2025
- The EU AI Act defines an AI system as a machine-based system that, for explicit or implicit objectives, infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions.EUR-Lex, Official Journal of the European Union, 2024
- The FCC ruled that calls using AI-generated voices need the prior express consent of the called party, unless an emergency purpose or an exemption applies.Federal Communications Commission, 2024
- The FCC's 2024 proposal states that TCPA requirements do not extend to technologies used to answer inbound calls.Federal Communications Commission, 2024
- The FCC's 2024 proposal would require callers to disclose to consumers when they receive an AI-generated call.Federal Communications Commission, 2024
- The Home Assistant Google Assistant integration lets you control your Home Assistant devices through Google Assistant on a phone, tablet or Google Home device.Home Assistant, 2026
- 334,000 pairs of N-best hypotheses and transcriptionsThe HyPoradise benchmark contains more than 334,000 pairs of recognizer N-best hypotheses and their correct transcriptions, for training language models to correct speech recognition errors.Chen Chen et al., arXiv (NeurIPS 2023 dataset track), 2023
- 334,000 pairsThe HyPoradise benchmark for language-model correction of speech recognition contains more than 334,000 pairs of N-best hypotheses and accurate transcriptions.Chen Chen et al., arXiv (NeurIPS 2023 dataset track), 2023
- The Moshi paper says current spoken dialogue systems rely on pipelines of independent components, and it casts spoken dialogue as speech-to-speech generation instead.Alexandre Défossez and and colleagues at Kyutai, arXiv (Kyutai), 2024
- The Moshi paper says that the complexity of current spoken dialogue pipelines induces a latency of several seconds between interactions.Alexandre Défossez and and colleagues at Kyutai, arXiv (Kyutai), 2024
- The Moshi paper says that when text is the intermediate step in dialogue, non-linguistic information that changes meaning is lost in the interaction.Alexandre Défossez and and colleagues at Kyutai, arXiv (Kyutai), 2024
- The original Piper text-to-speech repository was archived by its owner on October 6, 2025, and is read-only, with development moved to the OHF-Voice piper1-gpl repository.Rhasspy project contributors, GitHub, n.d.
- The participant in the 2023 Metzger study could not speak or make speech sounds, and could not type, because of weakness in her orofacial, vocal, arm and hand muscles.Sean L. Metzger et al., Nature (open full text via PubMed Central), 2023
- 253 channelsThe participant in the 2023 Metzger study was implanted with a 253-channel high-density electrocorticography (ECoG) array, placed on the surface of the brain.Sean L. Metzger et al., Nature (open full text via PubMed Central), 2023
- The PocketSphinx project says its algorithms and models are quite old, dating back to the 1970s in some cases, and that they remain useful because they are compact and efficient.CMU Sphinx project, n.d.
- 40.32 US dollars per monthThe same 2,520 monthly minutes cost about 40.32 US dollars a month at Google's standard rate of 0.016 dollars per minute.Calculated from other facts
- 6.30 US dollars per monthThe same 2,520 monthly minutes cost about 6.30 US dollars a month on AssemblyAI's Universal-2 model at 0.15 dollars per hour.Calculated from other facts
- 67 percent character error rateThe same 2025 preprint reports a character error rate of 67 percent when the decoder uses electroencephalography (EEG) recordings instead of MEG.Jarod Levy et al., arXiv (preprint), 2025
- The same 2025 study found that even with extensive background knowledge in the prompt, language models still struggled with the hardest domain-specific tasks.Qibang Liu et al., arXiv preprint 2502.14255, 2025
- The Speech-to-Phrase README says it is not a general purpose speech recognition system, because it answers which of the phrases it knows was said.OHF-Voice project contributors, GitHub, OHF-Voice organization, n.d.
- The Speech-to-Phrase README says the tool combines pre-defined sentence templates with the names of things from your Home Assistant to produce a file of the sentences it can recognize.OHF-Voice project contributors, GitHub, OHF-Voice organization, n.d.
- The Speech-to-Phrase repository holds the Kaldi-based version up to v1.4.3, and says development has moved to the OHF-Voice apps repository, where the rewrite is being built.OHF-Voice project contributors, GitHub, OHF-Voice organization, n.d.
- The US Copyright Office concludes that, on current technology, prompts alone do not give a user enough control over an AI output to make that user its author.United States Copyright Office, U.S. Copyright Office, 2025
- The US Copyright Office says that using AI tools to assist human creativity, rather than to stand in for it, does not affect whether the output can be protected by copyright.United States Copyright Office, U.S. Copyright Office, 2025
- The Vectara leaderboard asks each model to summarize every document using only the facts in that document, then scores the summary for factual consistency.Vectara, Vectara, GitHub, 2026
- The Voice Preview Edition has an internal speaker, an internal dual-microphone array and a 3.5mm audio output.Home Assistant project, Home Assistant, 2026
- 16 MB of flash storageThe Voice Preview Edition uses an ESP32-S3 system on a chip with 16 MB of flash storage and 8 MB of octal PSRAM, and an XMOS XU316 chip for audio processing.Home Assistant project, Home Assistant, 2026
- The VoiceAgentBench paper says existing speech benchmarks mostly test isolated skills such as transcription or question answering, and do not systematically test agentic behavior or adversarial robustness.Dhruv Jain et al., arXiv, 2025
- The VoiceAgentBench paper says large speech language models have enabled voice assistants that understand natural spoken queries and perform complex tasks.Dhruv Jain et al., arXiv, 2025
- 79 to 90 percent of failuresThe τ-Voice authors report that 79 to 90 percent of failures stem from agent behavior rather than from the speech input.Soham Ray et al., arXiv, 2026
- To connect Google Home to Gemini on an Android phone or tablet, you must be signed in to the Gemini mobile app with a personal Google Account.Google, n.d.
- 4 percent word error rateTraining on silent facial EMG reduced the transcription word error rate from 64 percent to 4 percent in one data condition of the 2020 study.David Gaddy and Dan Klein, Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020
- 300 wordsTurnitin raised the minimum length of a document its AI writing detector will evaluate from 150 words to 300 words.Annie Chechitelli, Turnitin, 2023
- 20 percentTurnitin says that in documents where its AI writing detector finds less than 20% AI writing, there is a higher incidence of false positives.Annie Chechitelli, Turnitin, 2023
- 0.0140 US dollars per minuteTwilio's US pricing page lists $0.0140 per minute for calls made to the United States and Canada, with pricing current as of August 2026.Twilio, 2026
- Under Article 50(1) of the EU AI Act, a provider must design AI systems that interact directly with people so that those people are informed they are interacting with an AI system.European Union, Regulation (EU) 2024/1689 (EU Artificial Intelligence Act), text mirror, 2024
- Under Article 50(1) of the EU AI Act, the duty to inform people that they are dealing with an AI does not apply where that is obvious to a reasonably well-informed, observant and circumspect person.European Union, Regulation (EU) 2024/1689 (EU Artificial Intelligence Act), text mirror, 2024
- US law bars calls to a residential line that use an artificial or prerecorded voice to deliver a message without the prior express consent of the called party, unless an exception applies.Cornell Law School Legal Information Institute, 1991
- 0.05 US dollars per minuteVapi charges $0.05 per minute for hosting on its pre-paid plan, and its speech and language models are billed at provider cost.Vapi, 2026
- 82 to 129 US dollars per month for 1,000 minutesVapi gives an example estimate of $82 to $129 per month for 1,000 minutes, using Deepgram transcription, OpenAI and ElevenLabs models, and telephony at $0.Vapi, 2026
- When Gemini launched on Android in February 2024, Google said starting a Routine in Gemini was not supported, and shortcuts and alarm-attached Routines on Android phones stopped working.Abner Li, 9to5Google, 2024
- Wispr Flow's Context Awareness help article says it skips context reading entirely in banking and financial apps.Wispr Flow Help Center, 2026
- Wispr Flow's Context Awareness help article says password field contents are never read or included in context.Wispr Flow Help Center, 2026
- Wispr Flow's Context Awareness help article says standard macOS password fields are excluded automatically, but custom or web-based password fields may be read like normal text fields.Wispr Flow Help Center, 2026
- Wispr Flow's current help article says its Context Awareness feature works from text and does not take screenshots.Wispr Flow Help Center, 2026
- Wispr Flow's help article says that its Context Awareness feature is on by default.Wispr Flow Help Center, 2026
- Wispr Flow's privacy page says its context awareness setting reads limited text from the app you are dictating into, so that names and jargon come out right.Wispr Flow (vendor website), 2026
- Wispr Flow's privacy policy says that if you enable its optional Context Awareness feature, it may collect limited, relevant content from the app in use, such as the text on the screen.Wispr, 2026
- 19.2 percent accuracy after eight editsWith eight adversarially chosen character edits, Mistral-7B-Instruct's accuracy on the GSM8K math benchmark fell to 19.2 percent.Esther Gan et al., arXiv preprint 2411.05345, 2024
- Yang and colleagues study whether large language models can act as post-processors for speech recognition, performing rescoring and error correction on the output of a recognizer.Chao-Han Huck Yang et al., IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) 2023, 2023
- You need a Matter hub inside your home to set up, control and automate Matter devices while you are away from home.Google, n.d.