Speech to Text With Translation on Phones and Computers

Speaking in one language and getting text in another usually happens in two steps. First a recognizer turns the sound into a transcript in the language you spoke. Then a translation model turns that transcript into the language you want. Many tools that do this follow the same design, and that design explains both what works and what goes wrong. This guide covers the three ways to do it on phones and computers, which apps offer each one, what limits they carry, how accurate the translation can be, and when the whole process stays on your device.

Speech translation is two steps: recognize the speech, then translate the text

The first step is recognition, and the second is translation. The design is called a cascade: a speech recognizer runs first, and its transcript goes into a separate text translator. Single-model systems such as SeamlessM4T are compared against those cascades in their paper ¹.

Developer services show the two steps clearly. Microsoft's Azure Speech service supports "real-time, multi-language speech to speech and speech to text translation of audio streams" ². Microsoft says interim transcription and translation results come back as speech is detected, before the final result ³. A live translation can therefore show partial text that is not final yet.

Not every model covers every direction. OpenAI's Whisper can translate speech into English, and OpenAI's speech-to-text guide names whisper-1 as the model for translating a complete recording into English ⁴. The turbo version of Whisper is not trained for translation, so translating another language into English needs one of the multilingual models instead ⁵. If you need a language other than English as the output, check the tool before you rely on it.

Translator apps are built for conversations: Google Translate, Microsoft Translator and Apple Translate

Three phone apps handle speech translation directly. Each one takes what you say, shows text on screen, and can play the translation aloud. They differ in how they handle offline use and turn-taking, and those differences matter more than the logos.

Google Translate puts speech input on its main screen. Google's help says that if the Speak button is disabled, the spoken language can't be translated ⁷. That is the first thing to check when voice is missing. Google's offline help says you can download languages onto your device to translate them without an internet connection ⁸. The page does not say whether speech works offline, so treat speech as online unless the app shows otherwise. For the step-by-step fix when voice is missing, read why Google Translate voice is not working, and for the basic steps read how to use voice input in Google Translate.

Microsoft Translator is strict about turns. Its Android FAQ says the maximum duration for each speaker in auto-mode speech translation is 20 seconds ⁶. Speak in short turns, and expect to pause between them. The same FAQ says offline packs work for text and camera translation only, and do not work in voice or conversation mode ⁹. Microsoft also tells users to check its Languages page to see whether a language supports speech at all ¹⁰. Do that check before a trip, not at the airport.

Apple Translate on iPhone uses Conversation mode. You can type a phrase or tap the microphone and speak, and the screen shows translated text bubbles from both sides. Apple says Conversation works for downloaded languages without an internet connection, or when you use on-device translation ¹¹. Its Auto Translate option removes the tap for each turn: "iPhone automatically detects when you start speaking and when you stop" ¹². Apple's guide is the only one of the three that states a speech mode working without a connection, and only for downloaded languages. Check that your languages are downloaded before you need them.

Which app to pick depends on the situation. For a face-to-face conversation on a phone, Apple's Conversation view with downloaded languages has the clearest offline path in the guides we checked. For a short exchange on Android, Google's Speak button is the simplest to find. For a long talk, the 20-second turn limit in Microsoft's Android app ⁶ shows that these apps are built for short turns. A recording you can play back is better handled by transcribing it first, which is covered in how to translate audio and video to English text.

Translating keyboards put translated text straight into any app

A translating keyboard is a different tool. It does not listen. It takes text you have entered and offers a translation that you insert into the message you are writing. On Android, Google's Gboard help describes the steps: open the features menu at the top of the keyboard, tap Translate, pick the language to translate from, pick the language to translate into, and enter your text ¹³. You then see a preview as you type and tap Select to use it.

That design has two consequences. First, because the Translate option sits on the keyboard itself, it is not tied to a single app. Second, the Translate step is built around typed text. Voice typing on the same keyboard is recognition, not translation ¹⁴. We did not find a Google source that says Gboard's Translate step accepts spoken input, so do not assume it does. If you want to test a spoken-then-translated message on a keyboard, try it in a message you can delete, and check the result before you send it.

Voice input is also a privacy question on keyboards. Google's App Store listing for Gboard on iPhone says that voice input from the Gboard microphone is sent to Google for transcription ¹⁴. A recognizer that runs on the phone does not make the translator on the phone, so check where each step runs. The privacy section below covers this in more detail.

On a computer, dictate in your language first and translate the text second

On a Mac or a Windows PC, the method that works with built-in tools takes two steps in sequence. You dictate in the language you speak, check the transcript, and then copy it into a translator. The Apple and Microsoft dictation pages we checked describe typing what you say, not translating it, so the translation step is a separate app or web page.

On a Mac, you can check whether your general dictation is processed on the device in the Keyboard settings ¹⁵. On Windows, voice typing needs an internet connection, a working microphone and the cursor in a text box before it will type what you say ¹⁶.

A practical workflow for a computer looks like this:

1. Dictate in the source language with the built-in dictation on your platform, and speak in full sentences.
2. Read the transcript and fix recognition errors before you translate. A wrong word in the transcript becomes a wrong word in the translation.
3. Copy the corrected text into a translator, such as a browser-based translation page or a translator app.
4. Check anything important with a speaker of the target language, especially names, numbers, dates and instructions.

The second step is the one people skip, and it is the one that saves the most trouble. The next section explains why a small recognition error carries into the translation.

Errors compound: a misheard word becomes a mistranslated word

The 2022 study by Tran and colleagues describes error propagation from the recognizer as a fundamental challenge that still remains in cascaded speech translation ¹⁷. The translation model sees only the transcript, so it translates a misheard word as faithfully as a correct one.

The size of the loss is measurable. A 2026 study on Vietnamese speech to English text found that translating the recognizer's output instead of clean transcripts cost between 6.79 and 10.64 BLEU points on the FLEURS benchmark ¹⁸. The same authors found that most substitution errors come from sound-alike words, and that these phonetic errors significantly degrade translation quality ¹⁹. So sound-alike errors are the ones this research singles out.

Recognition accuracy also varies by language, and the accuracy question gathers the numbers in one place. In the Whisper paper's tests on the FLEURS benchmark, the large-v2 model's word error rate was 4.2 percent for English and 3.0 percent for Spanish, but 21.5 percent for Hindi and 39.3 percent for Swahili ^{20 21 22 23}. The chart compares six languages on the same test. French was 8.3 percent ²⁴ and German 4.5 percent ²⁵.

The chart measures recognition only, on read speech, which is easier than a phone call in a noisy street. Every number in it is a word error rate for one language, so it says nothing about translation quality directly. For that, the research uses BLEU, an automatic score that compares machine output with reference translations. The 2002 paper that introduced BLEU proposed it as a quick, inexpensive, language-independent method that correlates highly with human evaluation ²⁶. Researchers use it because it is fast, not because it is perfect.

The table shows the scale of each effect, although the measurements use different data, so compare them with care. In one comparison, a direct model gained 1.3 BLEU over cascades ¹. In another, recognition errors alone cost several BLEU points ¹⁸. So the first place to look when a translation is wrong is the transcript. If the transcript is right and the translation still reads oddly, the fault is in the translation step.

Live translation works best with short, clear sentences and common language pairs

Speech translation rewards short, clear turns. The tools reflect this. The Android app of Microsoft's Translator limits each speaker to 20 seconds in auto mode ⁶. The Whisper paper says its models are trained on 30-second audio chunks and cannot take longer audio as a single input ²⁸. Shorter turns give the recognizer less to get wrong and give the translator a complete sentence to work with.

Speak in complete sentences with a pause at the end. Avoid trailing off or starting a new thought in the middle of a clause. Name proper nouns, medicines and street addresses slowly, and check them on the screen before you accept the translation. Apple's Auto Translate decides when a turn starts and stops, not whether the words were right, so check names on screen.

Language pairs matter as much as sentence length. Whisper's translation is aimed at English ⁴, so translating from English into another language needs another tool. Meta's SeamlessM4T is a single model that supports speech to text translation and recognition for up to 100 languages ²⁹. A recognition count is not a translation count, so check each pair you need. Microsoft's own advice is to look up whether each language supports speech at all ¹⁰.

The error rates for English and Spanish in the chart are far lower than for Swahili, so common pairs usually recognize speech better. For an uncommon pair, test the complete path before you need it: speak a sentence, read the transcript, and read the translation.

Voice translation is private only when both steps run on your device

Privacy depends on the steps, not the app name. A translator can be on-device while the recognizer is in the cloud, or the other way round. Speech translation is private only when both steps run locally, and you have to check each one in its own settings.

The practical rule is simple. If either step needs a connection, assume the audio or the transcript leaves your device at that step. Apple's Translate in Conversation mode is the only translation option in this table that states an offline path for speech, and only for downloaded languages. The other combinations depend on which translator you paste into. For a deeper look at what a recognizer keeps and where it goes, read on-device speech recognition compared with cloud recognition and the guide on offline voice typing on phones.

For developers, the same split shows up in the APIs. Azure Speech offers speech to text translation as one service ². Google Cloud's Basic translation API is a text service, "quick, plug-and-play access to the standard Neural Machine Translation (NMT) model" ³⁰, so a Google pipeline with speech needs a recognition call and then a translation call. The details are in which APIs combine speech to text and translation.

Where to go next

For the full overview of dictation languages, accents and voice translation, start with the guide to dictation languages, then the overview of dictation languages on iPhone and Android. If you are changing the language on a device, see how to change the dictation language. For speech from more than one language, see multilingual dictation, and for how accents affect recognition, see speech recognition and accents. For how recognition accuracy is measured, see how accurate speech recognition is, and for the word error rate itself, see word error rate. If you want to understand why the translation of a sentence changes when a single word is misheard, the glossary entry on speech translation explains the two-step design in more detail.

Accuracy claims in this guide come from tests on specific benchmarks and specific app versions, so a number here is a starting point. Test the pair you need, with the app you will use, and read the transcript before you trust the translation.