Facts

Languages and accents: facts and numbers

337 facts, each traced to the exact words of its source.

A 1986 UN working paper recalls that a UN conference recommended adopting Pinyin as the international system for romanizing Chinese geographical names.Sweden, United Nations Group of Experts on Geographical Names, 1986
A 2016 Interspeech paper says that Modern Standard Arabic is the official language of spoken and written Arabic media, while colloquial Arabic is the set of regional spoken varieties.Sarah Al-Shareef and Thomas Hain, Interspeech 2016, 2016
A 2017 study of YouTube's automatic captions found lower accuracy for women and for speakers from Scotland.Rachael Tatman, Proceedings of the First ACL Workshop on Ethics in Natural Language Processing, 2017
14.2 percent mixed error rateA 2023 adaptation of Whisper reached a 14.2 percent mixed error rate on Mandarin-English code-switched speech, training 5.6 percent additional parameters.Bobbi Aditya et al., arXiv, 2023
A 2024 Urdu benchmark found that Seamless-large was best on read speech, while Whisper-large was best on conversational speech.Samee Arif et al., arXiv, 2024
A 2025 Hebrew diacritics preprint says that diacritics are typically omitted in modern Hebrew writing, and that readers rely on familiarity to infer the meaning.Yair Elboher and Yuval Pinter, arXiv, 2025
A 2025 review of code-switching speech recognition found that a small subset of languages receive the most attention and that there is a large disparity in the languages covered by datasets.Maha Tufail Agro et al., arXiv (accepted at LREC 2026), 2025
A 2025 study says that error rates for Arabic speech rise sharply on dialectal and accented speech, compared with Modern Standard Arabic.Bashar Talafha et al., arXiv, 2025
6.79 to 10.64 BLEU pointsA 2026 study found that on the Vietnamese to English subset of FLEURS, translating recognizer output instead of clean transcripts cost between 6.79 and 10.64 BLEU.Giang Son Nguyen et al., arXiv, 2026
6.3 pointsA 2026 study found that scoring orthographic variants as correct, using its OIWER method, lowers error rates for Indian-language speech recognition by an average of 6.3 points.Kaushal Santosh Bhogale et al., arXiv preprint (AI4Bharat, IIT Madras), 2026
A 2026 study of open-source speech recognition models on Indian speech analyzes Hindi, Bengali, Kannada, Telugu and Marathi.Agneedh Basu et al., arXiv preprint, 2026
A December 2020 archived copy of Google's Docs help page lists Hindi among the languages and accents that voice typing works in.Google Docs Editors Help (archived by the University of North Texas Web Archive), 2020
A review chapter on pronunciation training says that speech recognition systems, such as those on Siri or Alexa, treat a learner's mispronunciations as noise, because their goal is to transcribe the sounds the speaker intended.Waris Quamer et al., The Encyclopedia of Applied Linguistics, 2nd edition (in press), author preprint, 2024
39 bits per secondA study of 17 languages found that they encode similar information rates, about 39 bits per second, despite large differences in speech rate and in information per syllable.Christophe Coupé et al., Science Advances, 2019
0.83 squared correlationAcross languages, the log of Whisper's word error rate has a squared correlation of 0.83 with the log of the amount of training data for that language.Alec Radford et al., OpenAI (arXiv), 2022
120 million parametersAI4Bharat's Hindi IndicConformer model is based on a Conformer-Large architecture with 120 million parameters and a hybrid CTC-RNNT decoder.AI4Bharat, AIKosh, the Government of India AI models portal, n.d.
AIKosh lists AI4Bharat's Hindi IndicConformer model under the MIT license, hosted by AI4Bharat.AI4Bharat, AIKosh, the Government of India AI models portal, n.d.
Amazon Transcribe lists two Arabic entries, Arabic (Gulf) with the code ar-AE and Arabic (Modern Standard) with the code ar-SA.Amazon Web Services, n.d.
Apple says Conversation works for downloaded languages without an internet connection, or when you use on-device translation.Apple, 2026
Apple says dictation on Apple TV may not be available in all countries or regions.Apple Support, 2026
Apple says that on a Mac, pressing and holding a letter key opens a menu of the accent marks you can enter with that letter.Apple, 2026
Apple says that with Auto Translate on, the iPhone detects when a speaker starts and stops, so you do not tap the microphone for each turn.Apple, 2026
Apple says words are translated as you type or speak in the Translate app's Conversation mode.Apple, 2026
Apple's auto-punctuation list for the Mac, which adds commas and periods as you dictate, includes German (Switzerland).Apple, Apple, 2026
Apple's Dictation list for the Mac includes both Portuguese (Brazil) and Portuguese (Portugal).Apple, Apple, 2026
Apple's feature availability page lists Japanese (Japan) and Korean (Korea) for Auto-Punctuation in Mac Dictation, alongside Chinese (Simplified), Chinese (Traditional) and Cantonese.Apple, Apple, 2026
Apple's feature availability page lists Japanese (Japan) and Korean (Korea) under On-Device and Modeless Dictation on the Mac, a separate list from general Dictation.Apple, Apple, 2026
Apple's feature availability page says that the on-device list of Dictation languages on the Mac requires a download of speech models.Apple, Apple, 2026
Apple's French Mac help says that Dictée, the French name for Dictation, adds some punctuation marks automatically as you dictate in supported languages.Apple Inc. (Aide Mac, French), 2026
Apple's French Mac table maps the spoken phrase 'deux-points' to the colon (:).Apple Inc. (Aide Mac, French), 2026
Apple's French Mac table maps the spoken phrase 'point d'interrogation' to the question mark (?) and 'guillemet ouvrant' to the opening guillemet («), with 'guillemet fermant' for the closing one (»).Apple Inc. (Aide Mac, French), 2026
Apple's German Mac help says that in supported languages the Dictation feature inserts certain punctuation marks automatically as you dictate.Apple Inc. (Mac-Hilfe, German), 2026
Apple's German Mac table maps the spoken phrase 'Anführungszeichen unten' to the low opening quotation mark („) and 'Anführungszeichen oben' to the high closing mark (“).Apple Inc. (Mac-Hilfe, German), 2026
Apple's German Mac table maps the spoken word 'Fragezeichen' to the question mark (?) and 'Ausrufezeichen' to the exclamation mark (!).Apple Inc. (Mac-Hilfe, German), 2026
Apple's German Mac table maps the spoken words 'Doppelpunkt', 'Komma' and 'Punkt' to the colon (:), the comma (,) and the full stop (.).Apple Inc. (Mac-Hilfe, German), 2026
Apple's iOS and iPadOS feature availability list for Dictation includes Arabic (Saudi Arabia) and Arabic (United Arab Emirates).Apple, 2026
Apple's Japanese Mac command table maps the spoken word カンマ (comma) to a half-width comma (,) and the spoken word ピリオド to a half-width period (.).Apple Inc. (Macユーザガイド, Japanese), 2026
18 variantsApple's Mac Dictation list has 18 language variants with automatic punctuation outside English and Spanish.Calculated from other facts
Apple's Mac feature availability list for Dictation includes Arabic (Saudi Arabia) and Arabic (United Arab Emirates).Apple, Apple, 2026
Apple's Mac feature availability list for Dictation includes Hebrew (Israel).Apple, Apple, 2026
Apple's Mac Help says that Dictation must be enabled before you can set its language in the Keyboard settings.Apple Support, n.d.
Apple's Mac Help says to click Keyboard in the sidebar of System Settings to reach the Dictation settings, and that you may need to scroll down to find it.Apple Support, n.d.
5 Spanish variantsApple's macOS feature availability page lists five Spanish variants for On-Device and Modeless Dictation on the Mac: Spanish (Chile), (Colombia), (Mexico), (Spain) and (United States).Apple, Apple, 2026
4 Spanish variantsApple's macOS feature availability page lists four Spanish variants for Auto-Punctuation in Mac Dictation: Spanish (Colombia), (Mexico), (Spain) and (United States). Spanish (Chile) is not in that list.Apple, Apple, 2026
Apple's main Dictation list for the Mac includes Dutch (Belgium) and Dutch (Netherlands).Apple, Apple, 2026
Apple's main Dictation list for the Mac includes French (Belgium), French (Canada), French (France) and French (Switzerland), and German (Austria), German (Germany) and German (Switzerland).Apple, Apple, 2026
Apple's main Dictation list for the Mac includes Italian (Italy) and Italian (Switzerland).Apple, Apple, 2026
Apple's on-device Dictation list for the Mac includes German (Switzerland) and French (Switzerland).Apple, Apple, 2026
Apple's on-device Dictation list for the Mac includes Portuguese (Brazil), but not Portuguese (Portugal).Apple, Apple, 2026
Apple's Simplified Chinese Mac table maps 逗号 to a full-width comma (，), 句号 to a full-width full stop (。), and 问号 to a full-width question mark (？).Apple Inc. (Mac 帮助, Simplified Chinese), 2026
Azure Speech returns one of the languages you supplied, even when none of them is spoken in the audio.Microsoft, Microsoft Learn, 2026
4 languagesAzure Speech's at-start language identification accepts up to four candidate languages.Microsoft, Microsoft Learn, 2026
10 languagesAzure Speech's continuous language identification accepts up to 10 candidate languages, while its at-start mode accepts up to four.Microsoft, Microsoft Learn, 2026
Azure's continuous language identification does not detect a change of language within a sentence. A Spanish sentence with English words inside it is not detected word by word.Microsoft, Microsoft Learn, 2026
BCP 47 says that a region subtag is typically used to indicate variations such as regional dialects or usage, or region-specific spelling conventions.Addison Phillips and Mark Davis, Internet Engineering Task Force (IETF), 2009
Deepgram's language table gives Arabic the base code ar and separate regional codes, including ar-EG for Egypt.Deepgram, n.d.
Dolev and colleagues conclude that Whisper can be used for transcribing Swiss German without adaptation, provided the desired output is Standard German.Eyal Liron Dolev et al., Proceedings of the 11th Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial 2024), 2024
11 languagesFor the 2025 Maha Kumbh, Bhashini's language translation supported a chatbot in 11 languages, including Hindi.Press Information Bureau, Government of India (Ministry of Electronics and Information Technology), 2025
5.65 to 43.03 percentage pointsFor the open source models in the test, the gap between the error rate on Svarah and on native English read speech from the Speech Accent Archive ranged from 5.65 to 43.03 percentage points.Tahir Javed et al., arXiv, 2023
Gallego found that dictogloss was better received by advanced-low learners, and that most learners found it useful and effective for learning.Muriel Gallego, Studies in Second Language Learning and Teaching, 2014
497 learnersGallego's 2014 dictogloss study had 497 learners at two proficiency levels.Muriel Gallego, Studies in Second Language Learning and Teaching, 2014
Google advises listing only the languages you need, because a shorter list makes it more likely that the service picks the correct language.Google, Google Cloud Documentation, 2026
Google Cloud says its Cloud Translation Basic API gives quick, plug-and-play access to the standard neural machine translation model.Google Cloud, 2026
Google Cloud Speech-to-Text lists Arabic (Egypt) with the code ar-EG for its chirp_3 model.Google Cloud, Google Cloud documentation, 2026
125+ languagesGoogle Cloud's Speech-to-Text product page says the service supports over 125 languages.Google Cloud, Google Cloud, 2026
Google says advanced voice typing in Gboard requires your primary Gboard language and your device language to be the same.Google, 2026
Google says automatic language detection is turned on by default in Gboard's advanced voice typing, on supported languages.Google, 2026
Google says supported languages for microphone translation vary by browser, with Chrome fully supported and limited support in Safari and Edge.Google Translate Help, 2026
Google says that changing the keyboard language in Gboard does not change the language settings of the Android device.Google, 2026
Google says that in some languages, you can hear the translation spoken aloud in Translate.Google Translate Help, 2026
Google says the Translate app can translate in near real-time while someone speaks a different language, for example a lecture or speech.Google Translate Help, 2026
Google says translation with a microphone on a computer does not automatically detect the language you speak, so you must choose it yourself.Google Translate Help, 2026
Google says you can download languages onto your device to translate them without an internet connection.Google, 2026
Google Translate speech input requires a device with a microphone.Google Translate Help, 2026
3 languagesGoogle's automatic language recognition accepts a list of up to three languages.Google, Google Cloud Documentation, 2026
Google's Cloud Translation language page lists Xhosa and Zulu in its Translation LLM supported languages, and says translations from any language to any other language in that list are supported.Google Cloud, 2026
14 regional Arabic entriesGoogle's Docs help says voice typing works in Arabic and in 14 regional Arabic entries, from Arabic (Algeria) to Arabic (United Arab Emirates).Google (Docs Editors Help), 2026
Google's Gboard help describes adding a language to Gboard from the Gboard settings, under Languages, with Add keyboard, and then choosing the language to turn on and a layout.Google, 2026
6 languagesGoogle's Gboard help lists six languages for advanced voice typing with automatic language switching: English, French, German, Italian, Japanese and Spanish.Google, 2026
Google's Gboard help says that talk-to-text does not work with all languages.Google, 2026
Google's help says that when the Speak button is disabled, the spoken language cannot be translated.Google, 2026
Google's Speech-to-Text overview says each alternative in the response contains the recognized text as a transcript. The overview does not describe translation.Google Cloud, 2026
Google's Speech-to-Text supported-languages page says developers should use only the language codes shown in its table, and that the codes are listed for each model and region.Google Cloud, Google Cloud documentation, 2026
In a 2002 controlled study of elementary EFL learners in Tehran, the group given frequent dictation made significantly larger listening gains than the control group.G. Reza Kiany and Ebrahim Shiramiry, TESL Canada Journal, 2002
46 to 71 percent relative word error reductionIn a 2007 study of foreign-accented English, the best combination of pronunciation modelling and acoustic adaptation cut word error rate by 46 to 71 percent relative.Ghazi Bouselmi et al., arXiv, 2007
7.84 syllables per secondIn a 2011 study that had speakers read translated texts aloud at a normal rate, Japanese speakers averaged 7.84 syllables per second, the fastest of the seven languages measured.François Pellegrino et al., Language (Linguistic Society of America), vol. 87, no. 3, 2011
44.5K hours of English training speechIn a 2021 test of a large English speech model trained on 44.5K hours of English speech, the authors found an accuracy bias that favored the accents most prevalent in the training corpus.Guillermo Cámbara et al., arXiv (2 pages, prepared for Phonetics and Phonology in Europe 2021), 2021
In a 2024 analysis of OpenAI's Whisper speech model, the authors found better performance on North American English accents than on British and Irish English accents.Calbert Graham and Nathan Roll, Cambridge Open Engage (preprint); published in JASA Express Letters, 2024, 2024
In a 2024 study of Arabic learners, a mnemonic method raised typing speed from 12 to 25 words per minute and accuracy from 55 to 85 percent.Nur Hanifansyah, Lisanan Arabiya: Jurnal Pendidikan Bahasa Arab, 2024
20.2 percent character error rateIn a 2024 study of Japanese speech recognition, the original Whisper-Base model had a character error rate of 20.2 percent on the study's Japanese test data.Mark Bajo et al., arXiv, 2024
32.7 percent character error rateIn a 2024 study, the untuned Whisper-Tiny model had a character error rate of 32.7 percent on Japanese, and fine-tuning on Japanese data lowered it to 14.7 percent.Mark Bajo et al., arXiv, 2024
86.10 to 93.06 percent word error rateIn a 2025 study of traditional Emirati Arabic, zero-shot Whisper Small had a word error rate of 93.06 percent and Whisper Medium had one of 86.10 percent.Maha AlBlooki et al., ACL Anthology (ICNLSP 2025 workshop paper), 2025
24 percentIn a 2026 test of YouTube's automatic Spanish captions, Argentinian Spanish had the highest overall word error rate of the dialects tested, at 24 percent.Iris Dania Jimenez and Christoph Kern, arXiv (Computation and Language, cs.CL), preprint 2602.24002, 2026
18 percentIn a 2026 test of YouTube's automatic Spanish captions, Mexican Spanish had an overall word error rate of 18 percent.Iris Dania Jimenez and Christoph Kern, arXiv (Computation and Language, cs.CL), preprint 2602.24002, 2026
23 words per minuteIn a two-minute transcription task, Japanese university students typed Japanese at 23 words per minute on desktop PCs.Chris Houser and Patricia Thornton, Kinjo Gakuin University (conference paper, copy hosted by Universidade Federal de Pernambuco), n.d.
In Apple's iOS list for on-device and modeless Dictation, Arabic appears without a country or region.Apple, 2026
In Apple's Spanish Mac command table, the spoken phrase 'signo de interrogación' produces a closing question mark (?), and the table has no command for the opening ¿.Apple Inc. (Mac Help, Spanish), 2026
23.0 percent word error rateIn Dolev and colleagues' tests, Whisper large-v3 used without any training on Swiss German had a word error rate of 23.0 percent on the STT4SG-350 test set.Eyal Liron Dolev et al., Proceedings of the 11th Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial 2024), 2024
In Google's Gboard help for Android, the translate steps are to pick the language to translate from, pick the language to translate into, and then enter your text.Google, 2026
In its January 2016 announcement, Google said the Devanagari alphabet keyboard let you type in Hindi using transliteration or handwriting input modes.Google, Google, The Keyword (blog.google), 2016
In January 2016, Google said that in English mode its keyboard let you type Hinglish and get suggestions for both Hinglish and English words.Google, Google, The Keyword (blog.google), 2016
36 percent of low-resource languagesIn Omnilingual ASR, only 36 percent of the languages with less than 10 hours of training audio meet the character error rate threshold of 10 percent.Omnilingual ASR team, arXiv, 2025
5 wordsIn running dictation, the runner is told to try to remember about five words at a time, and this can vary by level.Duncan Minett-Westwood, The Language Teacher, JALT, 2010
2.1 percent out-of-vocabulary rateIn Shin and colleagues' 2013 Russian system, a sub-word language model cut the out-of-vocabulary rate from 4.8 percent to 2.1 percent.Evgeniy Shin et al., Proceedings of IWSLT 2013, 2013
In Spanish, question marks are double signs: an opening mark (¿) and a closing mark (?) that must be placed at the start and end of an interrogative sentence.Wikipedia contributors, Wikipedia, 2026
11 dictationsIn the 2002 Tehran study, the experimental group was given dictation 11 times during the term, in addition to the textbook listening exercises.G. Reza Kiany and Ebrahim Shiramiry, TESL Canada Journal, 2002
26 studentsIn the 2025 Arabic pronunciation study, 26 American college students were assigned to two groups.Elsayed Issa and Gus Hahn-Powell, Language Learning & Technology, 2025
In the 2025 Arabic study, the group that used the automatic pronunciation trainer scored significantly better on the posttest than the in-class control group.Elsayed Issa and Gus Hahn-Powell, Language Learning & Technology, 2025
In the 2025 Emirati Arabic study, fine-tuning the five speech models on the Emirati corpus improved both their word error rate and their character error rate.Maha AlBlooki et al., ACL Anthology (ICNLSP 2025 workshop paper), 2025
7.62 percent character error rateIn the AISHELL-1 baseline study, the best of three 2017 models, LFMMI, reached a character error rate of 7.62 percent on the test set of Mandarin read speech.Hui Bu et al., arXiv (Oriental COCOSDA 2017), 2017
In the EdAcc study, every English speech recognition model tested showed a drop in performance on speakers of Indian, Jamaican and Nigerian English.Ramon Sanabria et al., arXiv (accepted to IEEE ICASSP 2023), 2023
In the FHNW team's 2022 evaluation, their own Swiss German model had almost half the word error rate of the best commercial system.Yanick Schraner et al., arXiv (Computation and Language), 2022
In the LAHAJA study, recognition performance declined for speakers from North-East and South India, especially on named entities and specialized terms.Tahir Javed et al., arXiv preprint (AI4Bharat, IIT Madras), 2024
6.19 syllables per secondIn the same 2011 study, English speakers averaged 6.19 syllables per second when reading the translated texts aloud at a normal rate.François Pellegrino et al., Language (Linguistic Society of America), vol. 87, no. 3, 2011
In the same 2025 study of non-native English, the five recognition systems varied significantly in how they handled disfluencies such as filler words, repetitions and revisions.Michael McGuire, arXiv, 2025
17 words per minuteIn the same two-minute tasks, Japanese university students typed Japanese at 17 words per minute on cellphones, against 10 on onscreen QWERTY keyboards.Chris Houser and Patricia Thornton, Kinjo Gakuin University (conference paper, copy hosted by Universidade Federal de Pernambuco), n.d.
In the Svarah test, the recognizers performed poorly for speakers whose first language is Assamese, Bodo or Nepali, which the authors say are low-resource languages from northeastern India.Tahir Javed et al., arXiv, 2023
4.3 times the Gemini error rateIn the Voxtral paper's FLEURS table, Whisper large-v3's Hindi word error rate was about 4.3 times that of Gemini 2.5 Flash.Calculated from other facts
Japanese government guidance on public writing (2022) says a single document must use one comma style throughout.Agency for Cultural Affairs, Japan (Bunka-cho), 2022
Krashen's input hypothesis says learners progress when they understand input that is slightly more advanced than their current level, which he called i+1.Wikipedia contributors, Wikipedia, n.d.
Lingua's English dictation page says the speaker reads each sentence twice, first at normal speed and then at a slower speed for easier writing.Lingua.com, n.d.
1,107 languagesMeta's Massively Multilingual Speech project built a single multilingual speech recognition model for 1,107 languages.Vineel Pratap et al., Meta AI (arXiv), 2023
58 percent relative word error rate reductionMeta's MMS model reduces Whisper's word error rate by a relative 58 percent on the 54 languages that both models support in FLEURS, while supporting over 11 times as many languages.Vineel Pratap et al., arXiv, 2023
78 percent of 1,570 evaluated languagesMeta's Omnilingual ASR paper reports that its 7B model meets the character error rate threshold for 78% of the 1,570 languages it was evaluated on.Omnilingual ASR team, arXiv, 2025
100 languagesMeta's SeamlessM4T is a single model that supports speech-to-text translation and speech recognition for up to 100 languages.Seamless Communication, Meta AI (arXiv), 2023
Microsoft defines an input method editor (IME) as a software component that supports text input for characters in languages that cannot be represented easily on a standard QWERTY keyboard, such as various East Asian languages.Microsoft Learn, n.d.
Microsoft gives the Settings path for adding a voice typing language that differs from your Windows display language: Time & language, then Language & region.Microsoft, 2026
Microsoft says interim transcription and translation results are returned as speech is detected, before the final results.Microsoft, 2026
Microsoft says its offline language packs work for text and camera translation only, and do not work in voice or conversation mode.Microsoft, 2026
Microsoft says that deleting your voice recording for Voice Isolation in Voice access removes your voice signature from the device.Microsoft, Microsoft Support, 2026
Microsoft says that on a hardware keyboard you turn on voice typing in Windows 11 by pressing the Windows key and H.Microsoft, 2026
Microsoft says that to switch voice typing languages on Windows you change the input language you use.Microsoft, 2026
Microsoft says that to switch Windows 11 voice typing languages, you change the input language you use.Microsoft, 2026
Microsoft says the preview languages in Dictate for Word may have lower accuracy and limited punctuation support than the main languages.Microsoft, 2026
Microsoft says the Speech service supports real-time, multi-language speech to speech and speech to text translation of audio streams.Microsoft, 2026
Microsoft says you are charged for the speech to text transcription and for the text translation for each target language.Microsoft, 2026
Microsoft says you can use a voice typing language that is different from the language you chose for Windows.Microsoft Support, 2026
Microsoft tells users to check the Languages page to see whether a language supports speech functionality.Microsoft, 2026
Microsoft's Arabic tables for Dictate in Word map the spoken phrase فاصلة (comma) to the Arabic comma (،) and علامة استفهام (question mark) to the Arabic question mark (؟).Microsoft, 2026
Microsoft's Azure Speech table of speech to text locales includes Hebrew (Israel), with the locale code he-IL.Microsoft, Microsoft Learn, 2026
Microsoft's Azure Speech table of speech to text locales includes Persian (Iran), with the locale code fa-IR.Microsoft, Microsoft Learn, 2026
Microsoft's Azure Speech table of speech to text locales lists separate Arabic locales for Egypt (ar-EG) and Morocco (ar-MA).Microsoft, Microsoft Learn, 2026
15 languagesMicrosoft's Dictate in Word article lists 15 main supported languages and locales, from Chinese (China) to Spanish (Mexico), with a separate list of preview languages.Microsoft, 2026
Microsoft's Dictate in Word lists spoken Spanish commands for the opening question mark and the closing question mark.Microsoft, 2026
Microsoft's French table for Dictate in Word maps the spoken phrase 'deux points' to a colon and 'guillemet gauche, guillemet ouvrant' to a left guillemet (« ).Microsoft, 2026
Microsoft's German table for Dictate in Word maps the spoken phrase for the opening quotation mark to « » (guillemets), not to the low German mark „.Microsoft, 2026
Microsoft's Hindi table for Dictate in Word maps the spoken phrase पूर्ण विराम (full stop) to the danda (।), the Devanagari sentence-ending mark.Microsoft, 2026
Microsoft's list of dictation commands for Word includes Hebrew (Israel).Microsoft, 2026
Microsoft's list of languages that support voice typing in Windows 11 includes French (Canada) and French (France) as two separate entries.Microsoft, 2026
Microsoft's list of languages that support voice typing in Windows 11 includes Hindi, Japanese, Korean, Marathi and Italian (Italy).Microsoft, 2026
Microsoft's list of languages that support voice typing in Windows 11 includes three Chinese entries: Simplified Chinese (China), Traditional Chinese (Hong Kong SAR) and Traditional Chinese (Taiwan).Microsoft Support, 2026
00000419 keyboard identifierMicrosoft's table of Windows keyboard layouts lists the Russian layout with the keyboard identifier 00000419.Microsoft Learn (Globalization), n.d.
20 secondsMicrosoft's Translator FAQ says the maximum duration for each speaker in auto-mode speech translation is 20 seconds.Microsoft, 2026
46 languagesMicrosoft's Windows 11 voice typing article lists 46 languages and locales that support voice typing, from Bulgarian to Vietnamese.Microsoft, 2026
Microsoft's Windows 11 voice typing page lists two Spanish variants, Spanish (Mexico) and Spanish (Spain), among its supported voice typing languages.Microsoft Support, 2026
Microsoft's Word dictation page warns that its Preview languages may have lower accuracy or limited punctuation support.Microsoft, 2026
2.50 US dollars per hourMicrosoft's worked example prices speech translation at a list price of $2.50 per hour, covering up to two target languages.Microsoft, 2026
Milde and Köhn say that compounds unknown to the language model are quite frequent in German, because German is a very productive language.Benjamin Milde and Arne Köhn, arXiv (Computation and Language), 2018
Milde and Köhn say that when the recognizer does not know a German compound, the error is counted as at least two errors.Benjamin Milde and Arne Köhn, arXiv (Computation and Language), 2018
Nemoto and colleagues studied the French homophones et (and) and est (is), which speech recognizers separate mainly with language model weights.Rena Nemoto et al., Proceedings of LREC 2008, 2008
117,000 hoursOf the 680,000 hours of audio used to train Whisper, 117,000 hours cover 96 languages other than English.Alec Radford et al., OpenAI (arXiv), 2022
117,000 hoursOf Whisper's 680,000 hours of training audio, 117,000 hours cover 96 languages other than English.Alec Radford et al., OpenAI (arXiv), 2022
On a computer, Google says to give your browser permission to use the microphone and to check the microphone settings in the browser.Google Translate Help, 2026
On an iPhone or iPad, Google says the transcription feature needs microphone permission for Google Translate turned on in the device settings.Google Translate Help, 2026
On an iPhone, Google says that if the Dictation button is unavailable, the spoken language cannot be translated.Google Translate Help, 2026
On Android, you tap Speak and wait for the prompt 'Speak now' before you say what you want translated.Google, 2026
On Apple TV 4K and Apple TV HD, Siri uses the language set up on the device, and you change it in Settings: General, then Siri, then Language.Apple Support, 2026
On Apple TV, the language button on the on-screen keyboard appears only if you have added additional languages in Settings.Apple Support, 2026
On Apple TV, you turn Siri dictation on or off in Settings, under General, then Dictation.Apple Support, 2026
0.054 mean match error rateOn read non-native English from the L2-ARCTIC corpus, Whisper reached a mean match error rate of 0.054 and AssemblyAI 0.056, which the author describes as approaching human-level accuracy.Michael McGuire, arXiv, 2025
0.063 mean match error rateOn spontaneous non-native English from the L2-ARCTIC corpus, RevAI had the best mean match error rate of the five systems tested, at 0.063.Michael McGuire, arXiv, 2025
7.2 percent word error rateOn Svarah, a 2023 benchmark of English from Indian speakers, the large version of OpenAI's Whisper had a word error rate of 7.2 percent in the Svarah column.Tahir Javed et al., arXiv, 2023
61 percent word error rateOn the 1996/97 NIST CallHome task for Egyptian Colloquial Arabic, the best word error rate obtained was 61 percent.Katrin Kirchhoff et al., IEEE Workshop on Automatic Speech Recognition and Understanding, 2003
80.3 percent accuracyOn the 82 languages that Whisper's training data covers, the best Whisper model identified the spoken language with 80.3% accuracy on the FLEURS benchmark.Alec Radford et al., OpenAI (arXiv), 2022
39.4 percent word error rateOn the Common Voice 9 test set, the Whisper large-v2 model had a word error rate of 39.4 percent for Persian.Alec Radford et al., OpenAI (arXiv), 2022
53.8 percent word error rateOn the Common Voice 9 test set, the Whisper large-v2 model had a word error rate of 53.8 percent for Arabic.Alec Radford et al., OpenAI (arXiv), 2022
26.1 BLEU pointsOn the CoVoST 2 test of speech translated into English, the Whisper large-v2 model scored 26.1 BLEU for Japanese speech translated into English.Alec Radford et al., OpenAI (arXiv), 2022
25.2 BLEU pointsOn the CoVoST 2 test of speech translated into English, zero-shot Whisper scored 25.2 BLEU for low-resource languages.Alec Radford et al., OpenAI (arXiv), 2022
36.2 BLEU pointsOn the CoVoST 2 test of speech translated into English, zero-shot Whisper scored 36.2 BLEU for high-resource languages.Alec Radford et al., OpenAI (arXiv), 2022
31.11 percent word error rateOn the CS-Dialogue Mandarin-English test set, Whisper Large-V2 had a 31.11 percent word error rate and a 15.29 percent mixed error rate, the highest error rates among the pre-trained models the authors tested.Jiaming Zhou et al., arXiv, 2025
19.7 percent word error rateOn the EdAcc accent test set of conversations between friends, the best-performing English speech recognition model, trained on 680 thousand hours of transcribed data, averaged a word error rate of 19.7 percent, against 2.7 percent on US English clean read speech.Ramon Sanabria et al., arXiv (accepted to IEEE ICASSP 2023), 2023
37.0 percent character error rateOn the Fleurs benchmark, the smallest Whisper model, tiny, had a character error rate of 37.0 percent for Japanese, against 5.3 percent for large-v2 on the same test.Alec Radford et al., OpenAI (arXiv), 2022
5.3 percent character error rateOn the Fleurs benchmark, the Whisper large-v2 model had a character error rate of 5.3 percent for Japanese, because the paper scores Japanese by characters rather than words.Alec Radford et al., OpenAI (arXiv), 2022
14.3 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 14.3 percent for Korean.Alec Radford et al., OpenAI (arXiv), 2022
16.0 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 16.0 percent for Arabic.Alec Radford et al., OpenAI (arXiv), 2022
17.5 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 17.5 percent for Tamil.Alec Radford et al., OpenAI (arXiv), 2022
21.5 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 21.5 percent for Hindi.Alec Radford et al., OpenAI (arXiv), 2022
22.6 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 22.6 percent for Urdu.Alec Radford et al., OpenAI (arXiv), 2022
27.1 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 27.1 percent for Hebrew.Alec Radford et al., OpenAI (arXiv), 2022
3.0 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 3.0 percent for Spanish.Alec Radford et al., OpenAI (arXiv), 2022
32.9 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 32.9 percent for Persian.Alec Radford et al., OpenAI (arXiv), 2022
39.3 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 39.3 percent for Swahili.Alec Radford et al., OpenAI (arXiv), 2022
4.2 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 4.2 percent for English.Alec Radford et al., OpenAI (arXiv), 2022
4.5 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 4.5 percent for German.Alec Radford et al., OpenAI (arXiv), 2022
5.3 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 5.3 percent for Japanese.Alec Radford et al., OpenAI (arXiv), 2022
5.6 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 5.6 percent for Russian.Alec Radford et al., OpenAI (arXiv), 2022
8.3 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 8.3 percent for French.Alec Radford et al., OpenAI (arXiv), 2022
94.8 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 94.8 percent for Yoruba.Alec Radford et al., OpenAI (arXiv), 2022
99.0 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 99.0 percent for Telugu.Alec Radford et al., OpenAI (arXiv), 2022
3.4 times the English error rateOn the Fleurs benchmark, the Whisper large-v2 word error rate in Korean was about 3.4 times its rate in English.Calculated from other facts
140.3 percent word error rateOn the FLEURS benchmark, Whisper large-v2 had a word error rate of 140.3 percent for Amharic, meaning more errors than words in the reference.Alec Radford et al., OpenAI (arXiv), 2022
88.9 percent word error rateOn the FLEURS benchmark, Whisper large-v2 had a word error rate of 88.9 percent for Hausa.Alec Radford et al., OpenAI (arXiv), 2022
1.8 timesOn the FLEURS benchmark, Whisper large-v2's French word error rate was about 1.8 times its German rate.Calculated from other facts
3.8 timesOn the Fleurs benchmark, Whisper large-v2's word error rate for Arabic was about 3.8 times its rate for English.Calculated from other facts
7.8 timesOn the Fleurs benchmark, Whisper large-v2's word error rate for Persian was about 7.8 times its rate for English.Calculated from other facts
1.4 times the Spanish error rateOn the Fleurs benchmark, Whisper large-v2's word error rate in English was about 1.4 times its rate in Spanish.Calculated from other facts
5.1 times the English error rateOn the FLEURS benchmark, Whisper large-v2's word error rate in Hindi was about 5.1 times its English rate.Calculated from other facts
5.1 times the English error rateOn the FLEURS benchmark, Whisper large-v2's word error rate in Hindi was about 5.1 times its rate in English.Calculated from other facts
4.6 times the Hindi error rateOn the FLEURS benchmark, Whisper large-v2's word error rate in Telugu was about 4.6 times its Hindi rate.Calculated from other facts
4.0 percent word error rateOn the FLEURS read-speech benchmark, Whisper large-v2 had a word error rate of 4.0 percent in Italian.Alec Radford et al., OpenAI (arXiv), 2022
4.3 percent word error rateOn the FLEURS read-speech benchmark, Whisper large-v2 had a word error rate of 4.3 percent in Portuguese.Alec Radford et al., OpenAI (arXiv), 2022
6.7 percent word error rateOn the FLEURS read-speech benchmark, Whisper large-v2 had a word error rate of 6.7 percent in Dutch.Alec Radford et al., OpenAI (arXiv), 2022
1.4 BLEU pointsOn the Fleurs read-speech test, Whisper large-v2 scored 1.4 BLEU when translating Yoruba speech into English.Alec Radford et al., OpenAI (arXiv), 2022
13.2 BLEU pointsOn the Fleurs read-speech test, Whisper large-v2 scored 13.2 BLEU when translating Bengali speech into English.Alec Radford et al., OpenAI (arXiv), 2022
17.2 BLEU pointsOn the Fleurs read-speech test, Whisper large-v2 scored 17.2 BLEU when translating Urdu speech into English.Alec Radford et al., OpenAI (arXiv), 2022
18.9 BLEU pointsOn the Fleurs read-speech test, Whisper large-v2 scored 18.9 BLEU when translating Japanese speech into English.Alec Radford et al., OpenAI (arXiv), 2022
22.0 BLEU pointsOn the Fleurs read-speech test, Whisper large-v2 scored 22.0 BLEU when translating Hindi speech into English.Alec Radford et al., OpenAI (arXiv), 2022
23.3 BLEU pointsOn the Fleurs read-speech test, Whisper large-v2 scored 23.3 BLEU when translating Spanish speech into English.Alec Radford et al., OpenAI (arXiv), 2022
34.6 BLEU pointsOn the Fleurs read-speech test, Whisper large-v2 scored 34.6 BLEU when translating German speech into English.Alec Radford et al., OpenAI (arXiv), 2022
7.2 BLEU pointsOn the Fleurs read-speech test, Whisper large-v2 scored 7.2 BLEU when translating Swahili speech into English.Alec Radford et al., OpenAI (arXiv), 2022
11.4 percent word error rateOn the FLEURS subset of the Vistaar Hindi benchmark, IndicWhisper had a word error rate of 11.4 percent.Kaushal Santosh Bhogale et al., Interspeech 2023 (arXiv:2305.15386), 2023
13.6 percent word error rateOn the Hindi subset of the Vistaar benchmark, IndicWhisper, which fine-tunes Whisper on Indian-language data, had an average word error rate of 13.6 percent across seven test sets.Kaushal Santosh Bhogale et al., Interspeech 2023 (arXiv:2305.15386), 2023
16.5 percent character error rateOn the third evaluation set of the Corpus of Spontaneous Japanese (CSJ eval3), Whisper large-v3 had a character error rate of 16.5 percent in the paper's comparison table.Yosuke Kashiwagi et al., arXiv (Interspeech 2025), 2025
39 percent mixed error rateOn two conversational code-switching corpora, Whisper's mixed error rate on code-switched utterances was about 39 percent.Sarina Meyer et al., arXiv, 2025
OpenAI says Whisper's performance varies widely depending on the language, and it publishes a per-language breakdown of error rates.OpenAI, GitHub (OpenAI repository), 2023
OpenAI's speech-to-text guide reserves its specialized models for cases such as subtitle formats and translation into English.OpenAI, OpenAI, 2026
OpenAI's translations endpoint supports translation into English only.OpenAI, OpenAI, 2026
OpenAI's Whisper README says the model's performance varies widely depending on the language, and it publishes a breakdown by language for its two largest models.OpenAI, GitHub (openai/whisper), 2026
30 to 50 percent relative increaseRecent evaluations, which the HiACC authors cite, show that ASR models have a 30 to 50 percent relative increase in word error rate on code-switched speech compared with monolingual input.Shruti Singh et al., Data in Brief (via PubMed Central), 2025
Samsung's US support page for the Interpreter app lists its supported languages, and Bengali and Urdu are not on that list.Samsung Electronics America, 2026
1.3 BLEU pointsSeamlessM4T improved into-English speech-to-text translation by 1.3 BLEU points over strong cascaded systems.Seamless Communication, Meta AI (arXiv), 2023
SeamlessM4T's language table marks Xhosa as zero-shot for speech translation into English.Seamless Communication, Meta AI (arXiv), 2023
SeamlessM4T's language table marks Zulu as a low-resource language for speech translation into English.Seamless Communication, Meta AI (arXiv), 2023
Shin and colleagues describe Russian as a challenging language for automatic speech recognition because of its rich morphology.Evgeniy Shin et al., Proceedings of IWSLT 2013, 2013
6 grammatical casesShin and colleagues write that Russian nouns can be declined in six cases, with singular and plural numbers and three grammatical genders.Evgeniy Shin et al., Proceedings of IWSLT 2013, 2013
55 percentThe 11 dictations in the 2002 Tehran study equal 55 percent of the term's 20 sessions, if each dictation fell in a separate session.Calculated from other facts
The 2002 paper that introduced BLEU proposed a quick, inexpensive, language-independent automatic method that correlates highly with human evaluation.Kishore Papineni et al., Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, 2002
The 2017 MGB-3 Arabic challenge emphasized dialectal Arabic and drew its data from Egyptian YouTube videos.Ahmed Ali et al., arXiv, 2017
The 2022 NAACL paper says that because most vowels in Hebrew are encoded as diacritics, the pronunciation of words is left underspecified.Elazar Gershuni and Yuval Pinter, Findings of the Association for Computational Linguistics: NAACL 2022, 2022
The 2023 Interspeech paper found that diacritizing inside the speech model outperformed diacritizing the text afterward, especially when the model was fine-tuned on manually diacritized transcripts.Hanan Aldarmaki and Ahmad Ghannam, Interspeech 2023, 2023
The 2023 Interspeech paper says that most people write and type Arabic text without diacritics.Hanan Aldarmaki and Ahmad Ghannam, Interspeech 2023, 2023
The 2023 Interspeech paper says that most speech corpora and Arabic ASR models are undiacritized, because diacritics are often omitted.Hanan Aldarmaki and Ahmad Ghannam, Interspeech 2023, 2023
The 2024 official German rules print direct speech with the low opening quotation mark („) and the high closing quotation mark (“).Rat für deutsche Rechtschreibung, 2024
The 2026 Vietnamese study found that most recognition substitution errors come from sound-alike confusions, and that these errors significantly degrade translation quality.Giang Son Nguyen et al., arXiv, 2026
The Arabic comma (U+060C, ARABIC COMMA) was added to Unicode in version 1.1, in 1993.codepoints.net (Unicode character data), n.d.
The authors attribute the missing tildes mainly to speech recognition techniques that ignore stress, the stressed syllable of a word.Constantino Bolaños Araya et al., Kañina, Universidad de Costa Rica, 2016
The authors found systematic differences in how accurately YouTube's Spanish captions handled different dialects, which they attribute to specific dialects rather than to chance.Iris Dania Jimenez and Christoph Kern, arXiv (Computation and Language, cs.CL), preprint 2602.24002, 2026
The authors found that speech recognition with Whisper significantly altered and reduced the speakers' original code-switching.Sarina Meyer et al., arXiv, 2025
The authors of a 2022 study call error propagation from the speech recognition system a fundamental challenge for cascaded speech translation that still remains.Viet Anh Khoa Tran et al., arXiv, 2022
The authors say that in Spanish, one of the most common errors of automatic speech recognition is the omission of the diacritical tilde.Constantino Bolaños Araya et al., Kañina, Universidad de Costa Rica, 2016
The authors say that multilingual models such as Whisper often struggle with the code-switching setting.Bobbi Aditya et al., arXiv, 2023
7.3 timesThe best model's average word error rate on EdAcc conversational speech was about 7.3 times its word error rate on US English clean read speech.Calculated from other facts
The CAMÕES authors state that existing resources for automatic speech recognition in Portuguese are mostly focused on Brazilian Portuguese, which leaves European Portuguese under-explored.Carlos Carvalho et al., arXiv, 2025
46 hoursThe CAMÕES benchmark includes a 46-hour test set of European Portuguese speech, covering multiple domains.Carlos Carvalho et al., arXiv, 2025
The Common Core grade 1 standard RF.1.3.g asks students to recognize and read grade-appropriate irregularly spelled words.National Governors Association Center for Best Practices and Council of Chief State School Officers, Common Core State Standards Initiative, 2010
The Common Core kindergarten foundational skills standard 3 asks children to know and apply phonics and word analysis skills in decoding words.National Governors Association Center for Best Practices and Council of Chief State School Officers, Common Core State Standards Initiative, 2010
The Common Core kindergarten language standard 2 asks children to show command of standard English capitalization, punctuation and spelling when they write.National Governors Association Center for Best Practices and Council of Chief State School Officers, Common Core State Standards Initiative, 2010
The daṇḍa, the vertical stroke used as a full stop in Hindi and Nepali, marks the end of a sentence or line.Wikipedia, n.d.
The Dragonfly documentation lists Japanese among the languages supported by Windows Speech Recognition as of 2016. The same page lists Dragon's languages separately and does not include Japanese in them.Dragonfly project, Dragonfly documentation (Read the Docs), n.d.
6 languagesThe Dragonfly documentation says Dragon Professional Individual version 15 supports English in several variants, plus Dutch, French, German, Italian and Spanish, and it cites Nuance's knowledgebase for other editions.Dragonfly project, Dragonfly documentation (Read the Docs), n.d.
श्रुतलेखनThe English Wiktionary entry for dictation gives श्रुतलेखन (śrutlekhan) as the Hindi translation of the noun.Wiktionary, n.d.
The faster-whisper transcribe method has a language_detection_segments option that sets how many segments are considered for language detection.SYSTRAN, GitHub, SYSTRAN/faster-whisper, 2026
The FLEURS paper says that Japanese and Korean transcripts may or may not contain spaces, depending on the transcriber, so the benchmark needed its own handling of spacing.Alexis Conneau et al., arXiv (IEEE Spoken Language Technology Workshop 2022), 2022
The HinKhoj English-to-Hindi dictionary gives लिखवाना, to have something written down, as the second Hindi meaning of the verb dictate, with the example of a lawyer dictating a letter to a secretary.HinKhoj Dictionary, n.d.
श्रुतलेखThe HinKhoj English-to-Hindi dictionary gives श्रुतलेख as the Hindi meaning of dictation, a shorter spelling of the word that Wiktionary lists as श्रुतलेखन.HinKhoj Dictionary, n.d.
The HinKhoj English-to-Hindi dictionary gives हुक्म देना, to give orders, as the first Hindi meaning of the verb dictate. That sense is different from writing down speech.HinKhoj Dictionary, n.d.
The Hong Kong handbook says a phonics dictation revises vocabulary the pupils have learnt and also has them spell less familiar words by applying phonics skills.Hong Kong Education Bureau, Hong Kong Education Bureau, English Language Education, n.d.
The Hong Kong handbook says dictation, like all English learning activities, should be contextualised so pupils can seek and process information.Hong Kong Education Bureau, Hong Kong Education Bureau, English Language Education, n.d.
The Hong Kong handbook says marks should not be deducted for repeated mistakes in a dictation, because repeating the same misspelled word does not show the same ability as several different errors.Hong Kong Education Bureau, Hong Kong Education Bureau, English Language Education, n.d.
22 languagesThe IndicVoices authors built IndicASR, which they describe as the first speech recognition model to support all 22 languages in the Eighth Schedule of the Indian Constitution.Tahir Javed et al., arXiv preprint (AI4Bharat, IIT Madras), 2024
7348 hours of speechThe IndicVoices dataset contains 7,348 hours of read, extempore and conversational speech from 16,237 speakers, covering 22 Indian languages.Tahir Javed et al., arXiv preprint (AI4Bharat, IIT Madras), 2024
10.31 percent character error rateThe KoSpeech baseline reached a character error rate of 10.31 percent on the KsponSpeech corpus of spontaneous Korean, using the acoustic model only.Soohwan Kim et al., arXiv, 2020
The LAHAJA authors evaluated existing open-source and commercial Hindi models and found their performance poor.Tahir Javed et al., arXiv preprint (AI4Bharat, IIT Madras), 2024
12.5 hours of Hindi audioThe LAHAJA Hindi benchmark contains 12.5 hours of Hindi audio from 132 speakers across 83 districts of India.Tahir Javed et al., arXiv preprint (AI4Bharat, IIT Madras), 2024
The LANGUAGES table in OpenAI's Whisper source code includes Arabic, with the code ar.OpenAI, GitHub, openai/whisper, 2026
The LANGUAGES table in OpenAI's Whisper source code includes Bengali, with the code bn, and Urdu, with the code ur.OpenAI, GitHub, openai/whisper, 2026
The LANGUAGES table in OpenAI's Whisper source code includes Bengali, with the code bn.OpenAI, GitHub, openai/whisper, 2026
The LANGUAGES table in OpenAI's Whisper source code includes Urdu, with the code ur.OpenAI, GitHub, openai/whisper, 2026
The MGB-3 Arabic challenge data totals 16 hours of video, split evenly across seven genres.Ahmed Ali et al., arXiv, 2017
about 100 languagesThe MMS authors say current speech technology is limited to about one hundred languages, a small fraction of the over 7,000 languages spoken in the world.Vineel Pratap et al., Meta AI (arXiv), 2023
99 languagesThe MMS paper says Whisper's model supports the transcription of 99 different languages.Vineel Pratap et al., arXiv, 2023
The Open Universal Arabic ASR Leaderboard is a continuous benchmark of open-source Arabic speech models across several multi-dialect datasets.Yingzhi Wang et al., arXiv, 2024
The Québec language office's spacing table gives a non-breaking space after the opening French guillemet (« ) and before the closing one (»).Office québécois de la langue française, n.d.
The Québec language office's spacing table shows the colon (deux-points) with a non-breaking space before it and a normal space after it.Office québécois de la langue française, n.d.
2.7 percent word error rateThe same best English speech recognition model, tested by the EdAcc authors on US English clean read speech, had a word error rate of 2.7 percent.Ramon Sanabria et al., arXiv (accepted to IEEE ICASSP 2023), 2023
The SeamlessM4T authors write that conventional speech-to-speech translation systems rely on cascaded systems that perform translation progressively.Seamless Communication, Meta AI (arXiv), 2023
The SeamlessM4T paper lists error propagation from speech recognition into the translation model as a known weakness of cascaded systems.Seamless Communication, Meta AI (arXiv), 2023
The SeamlessM4T paper reports a language as zero-shot for speech translation into English when it lacks the primary training data counted in its table.Seamless Communication, Meta AI (arXiv), 2023
1.3 BLEU pointsThe SeamlessM4T paper says that, compared with strong cascaded models, its speech to text translation into English is 1.3 BLEU points better.Seamless Communication, Meta AI (arXiv), 2023
343 hoursThe STT4SG-350 corpus holds 343 hours of Swiss German speech from all dialect regions.Michel Plüss et al., Proceedings of the 61st Annual Meeting of the ACL (Volume 2: Short Papers), 2023
The study's authors quote Roach (1999) that a language with simple syllables such as Japanese can fit more syllables into a second than English or Polish, so it will probably sound faster.François Pellegrino et al., Language (Linguistic Society of America), vol. 87, no. 3, 2011
The Svarah authors found that the two commercial systems they tested were outperformed by the open source Whisper and NVIDIA Conformer models on Indian English.Tahir Javed et al., arXiv, 2023
The Svarah authors say Indian speakers are very poorly represented in existing English speech recognition benchmarks such as LibriSpeech and Switchboard.Tahir Javed et al., arXiv, 2023
The Svarah authors say the medium and large versions of Whisper gave single-digit word error rates on Indian English, and note that their size makes them hard to deploy in India.Tahir Javed et al., arXiv, 2023
14.0 percent word error rateThe Swiss German corpus authors fine-tuned a Wav2Vec2 XLS-R model on their data and reached a word error rate of 14.0 percent on their test set, as Dolev and colleagues report.Eyal Liron Dolev et al., Proceedings of the 11th Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial 2024), 2024
20 sessionsThe term in the 2002 Tehran study consisted of 20 sessions.G. Reza Kiany and Ebrahim Shiramiry, TESL Canada Journal, 2002
The Unicode bidirectional algorithm calls the left-to-right and right-to-left bidirectional types strong types, and characters of those types strong directional characters.Unicode Consortium, Unicode Consortium, 2026
The Unicode Standard prescribes a memory representation order known as logical order, which is the order characters are stored in, separate from the order they are displayed in.Unicode Consortium, Unicode Consortium, 2026
The Unicode standard says that when right-to-left scripts use digits written from left to right, the text is actually bidirectional, a mixture of right-to-left and left-to-right text.Unicode Consortium, Unicode Consortium, 2026
12 languagesThe Vistaar benchmark collects publicly available Indian-language test sets across 12 languages, giving 59 computed word error rate values.Kaushal Santosh Bhogale et al., Interspeech 2023 (arXiv:2305.15386), 2023
16x increase in training data per halvingThe Whisper authors estimate that a language's word error rate halves for every 16-fold increase in its training data.Alec Radford et al., OpenAI (arXiv), 2022
The Whisper authors say their worst outliers are languages with unique scripts that are distantly related to the Indo-European languages that make up most of the training data.Alec Radford et al., OpenAI (arXiv), 2022
The Whisper paper names five languages whose text it scores by character error rate: Chinese, Japanese, Thai, Lao and Burmese. Korean is not on that list.Alec Radford et al., OpenAI (arXiv), 2022
29.1 BLEUThe Whisper paper reports a zero-shot score of 29.1 BLEU on the CoVoST2 speech to English translation benchmark, without training on CoVoST2 data.Alec Radford et al., OpenAI (arXiv), 2022
99 languagesThe Whisper paper says its model predicts the spoken language with one token for each language in its training set, and that there are 99 such tokens, so the training data covers 99 languages.Alec Radford et al., OpenAI (arXiv), 2022
The Whisper paper's authors found that most of the training data labeled as Welsh-to-English translation was actually English audio.Alec Radford et al., OpenAI (arXiv), 2022
The Whisper README says its default setting selects the turbo model, and that the default works well for transcribing English.OpenAI, GitHub (openai/whisper), 2026
The Whisper README says that the turbo model returns the original language when the translate task is set, and recommends the medium or large model for the best translation results.OpenAI, GitHub (openai/whisper), 2026
The Whisper README says the turbo model returns the original language even when the translate task is specified.OpenAI, GitHub (openai/whisper), 2026
The Whisper repository says its turbo model is not trained for translation, and that you should use a multilingual model (tiny, base, small, medium or large) to translate non-English speech into English.OpenAI, GitHub (OpenAI repository), 2023
Transliteration is the process of representing a word, phrase or text in a different script or writing system.Wikipedia, n.d.
Unicode encodes the daṇḍa as code point U+0964, named DEVANAGARI DANDA.Wikipedia, n.d.
Unicode says that in Hindi the word hal names the character that suppresses the inherent vowel, and halant names the consonant whose vowel is suppressed.Unicode Consortium, 2022
Unicode says the Brahmic scripts are abugidas in which most symbols stand for a consonant plus an inherent vowel, usually the sound a.Unicode Consortium, 2022
Unicode says the dependent vowels, written on a consonant, are the common way to show vowels other than the inherent one, and that they are called matras in Sanskrit.Unicode Consortium, 2022
Unicode's bidirectional algorithm changes a European number to an Arabic number when the nearest strong character before it is an Arabic letter.Unicode Consortium, Unicode Consortium, 2026
Unicode's bidirectional algorithm says that neutral characters, such as punctuation and spaces, generally take the direction of the surrounding text, and take the embedding direction when the two sides conflict.Unicode Consortium, Unicode Consortium, 2026
Unicode's DEVANAGARI DANDA (U+0964) is the Devanagari character that ends sentences, as the Hindi full stop.codepoints.net (Unicode character data), n.d.
Unicode's IDEOGRAPHIC FULL STOP (U+3002), in the CJK Symbols and Punctuation block, is the character used to end sentences in CJK text.codepoints.net (Unicode character data), n.d.
When a request lists several languages, Google's service transcribes the audio using one best-fit language from that list.Google, Google Cloud Documentation, 2026
30 secondsWhen no language is set, Whisper's transcribe function detects the spoken language from up to the first 30 seconds of audio.OpenAI, GitHub (openai/whisper), 2026
Whisper's language table has one code for German, de, and one for Portuguese, pt, with no separate Swiss German code and no separate Brazilian or European Portuguese code.OpenAI, GitHub, openai/whisper, 2026
20 languages without training dataWhisper's training data contains no speech for 20 of the 102 languages in the FLEURS benchmark, which caps its language identification accuracy at 80.4%.Alec Radford et al., OpenAI (arXiv), 2022
125,000 hoursWhisper's training data includes 125,000 hours of audio for translating speech from other languages into English.Alec Radford et al., OpenAI (arXiv), 2022
125,000 hours of X to English translation dataWhisper's training data includes 125,000 hours of translation data from speech in other languages into English.Alec Radford et al., OpenAI (arXiv), 2022
Wikipedia describes InScript as the decreed standard keyboard layout for Indian scripts.Wikipedia contributors, Wikipedia, The Free Encyclopedia, n.d.
Wikipedia says that alternative Russian phonetic layouts map Cyrillic letters to similar-sounding Latin letters on QWERTY and other layouts.Wikipedia, 2026
Wikipedia says that JCUKEN is the main keyboard layout for the Russian language on computers and typewriters.Wikipedia, 2026
1953 yearWikipedia says that the JIUKEN layout was the main Russian layout before it was replaced by JCUKEN in 1953.Wikipedia, 2026
Wikipedia says the Devanagari InScript bilingual layout has a common layout for all the Indian scripts, and that most Indic scripts share the same phonetic character order.Wikipedia contributors, Wikipedia, The Free Encyclopedia, n.d.
Wikipedia says the final schwa in its example word, a Sanskrit word, is deleted in both Hindi and Bengali.Wikipedia, n.d.
Windows 11 voice typing lists Dutch (Netherlands) as its Dutch entry, with no Belgian Dutch entry.Microsoft, 2026
Windows 11 voice typing lists German as one language with no regional variant, and it has no Swiss German entry.Microsoft, 2026
Windows 11 voice typing lists Italian (Italy) as its Italian entry.Microsoft, 2026
Windows 11 voice typing lists Portuguese (Brazil) and Portuguese (Portugal) as separate languages.Microsoft, 2026
Windows 11 voice typing lists Russian as one language, with no region in its name.Microsoft, 2026
YouTube's 2008 launch post says auto translate adds machine translation to the caption tracks that uploaders add to a video.The YouTube Team, YouTube Blog, 2008