Live French transcription, wherever people speak
French speech to text for voice agents serves a person who expects a reply right away, whether they are cooking next to a smart speaker, practicing in a tutor app or asking a website for help. Hear receives a live stream over a WebSocket and returns partial text in about 200 ms. PyAI Omni, the speech to speech agent, is built on Hear, so you can plug Hear into your own assistant or let Omni listen, reason and answer in a single session. The same feed gives live French transcription for subtitles or notes. The same engine also works as a French speech to text generator for the transcripts your team keeps after each session.
Key takeaways
- First partial back: a French partial arrives about 200 ms after speech starts.
- Accuracy, English measured: 1.59% WER clean and 4.76% on 8 kHz phone, accented or noisy audio, PyAI benchmarks at industry parity. No French figure is published.
- Channels: smart speakers, apps, web widgets and lines share one streaming endpoint.
- Cost: $0.001/min, billed per whole minute, and empty transcripts are free.
- Omni timing: end of turn near 196 ms, voice to voice around 390 ms median.
- Switching: whisper-1 compatible, so only the base URL changes.
How spoken French turns into a response
A kitchen speaker, a browser tab and a mobile app all follow the same path from sound to text to action.
Someone speaks French to a device, a web page, an app or a line, and the client captures the microphone.
PCM16 frames of roughly 20 ms travel over a WebSocket while the person is still speaking.
Hear answers with a first partial around 200 ms in and keeps sharpening the text.
Your agent, or Omni, sets the timer, answers the question or updates the list from the final words.
French speech situations Hear covers
French input varies by region, runs words together and mixes in English. Here is what to plan for in any channel.
France, Quebec, Belgium, West Africa
Accents and vocabulary shift a lot between Paris, Montréal, Brussels and Dakar. Test with recordings from your own users, since PyAI publishes no French score.
Liaison and elision
Phrases like les amis or j'arrive blur word edges at speed. Review a few fast samples to see how word boundaries come out.
Counting the French way
Quatre vingt dix and numbers read in pairs return as digits with numerals handling, so a timer or a price reaches your logic clean.
Anglicismes and brand names
French users say le week-end or name English apps. Up to 5 custom vocabulary terms keep those names spelled your way.
Kitchens and living rooms
Extractor fans, music and children add noise. Hear measures 4.76% WER on noisy and accented audio, a number measured on English.
Partial text and timing
Text flows before the request ends, and verbose JSON adds speaker turns and timestamps for logs and subtitles.
Accuracy and speed
Accuracy figures are PyAI's own English measurements at industry parity. Use them as a guide for French, then check samples from your own product. Omni timings were measured inside full Omni sessions.
French use cases for a listening assistant
French voice assistant speech to text belongs anywhere French speakers want to talk instead of type. Six use cases, all equal:
Smart speakers and home hubs
Turn spoken French into timers, shopping lists and device commands on a speaker or hub.
Voice tutors
Score a learner's spoken French against the expected sentence and show where the transcript differs.
Website help widgets
Let a visitor in Lyon or Montréal explain an issue aloud on a support page instead of filling a form.
Game characters
Feed a player's spoken lines to an in-game character so the story reacts to what was said.
Accessibility dictation
Help people who find typing hard write emails, forms and notes in French by speaking.
Customer support phone line
Answer a French support number, work out what the caller wants and route or resolve the request.
Formats, limits and output
| Input | French audio as PCM16 over a live WebSocket from a speaker, app, browser or line, or one file up to 25 MB. Any whisper-1 client connects as is. |
|---|---|
| Output | Partials while the user speaks, then the final transcript, as text, JSON or verbose JSON with timestamps and speaker turns. |
| Modes | Sync file call for short clips, live streaming for apps, devices and voice sessions, and background jobs for long recordings that run 92 to 247 times real time. |
| Tuning | Choose French, enable numerals handling for times and quantities, and add up to 5 custom vocabulary terms such as product or brand names. |
| Billing | $0.001/min of audio, rounded to the whole minute. An empty transcript costs nothing, and a busy engine returns a retry, not a charge. |
Plug Hear into a French assistant
Since Hear follows the OpenAI transcription format, a French assistant on whisper-1 moves across with a base URL edit. Docs: PyAI SDKs.
SDK
On your server, the SDK opens a stream for each French session and passes partial and final text to your assistant logic the instant Hear produces it.
MCP
With the MCP server, an agent can call Hear as a tool, for instance to transcribe a French voice memo before choosing what to do.
CLI
Run the CLI on a French recording or your own mic and watch partials print live while you tune vocabulary terms.
Where Hear fits
PyAI is voice infrastructure. Hear, Speak and Omni share one stack in one region, so a French session runs from the opening syllable to the reply without changing vendor. Hear is not a conversation designer and does not choose what your assistant replies. Results on PyAI benchmarks come from English tests, so trial Hear on real French speech first.
French voice agent questions
What latency does Hear give on French speech?
Around 200 ms to the first partial. Hear streams French text back roughly 200 ms after the user starts talking and updates it as the sentence continues. When Hear runs inside Omni, the end of a turn is caught about 196 ms after silence, and voice to voice lands near 390 ms median.
Does PyAI publish a French word error rate?
No. The Hear numbers PyAI publishes, 1.59% on clean audio and 4.76% on 8 kHz phone, accented or noisy audio, were measured on English and sit at industry parity. French teams should run recordings from their own users in France, Quebec or Africa and read the transcripts closely.
Can my existing French assistant use Hear?
Yes. Your service opens one WebSocket per session, streams PCM16 from the speaker, app or browser, and hands Hear's partial and final French text to your own reasoning and voice. If the assistant already calls whisper-1, changing the base URL is usually the only edit needed.
What is the price of French transcription with Hear?
Hear bills $0.001 per minute of audio, counted by the whole minute, with no platform fee, and a session that returns an empty transcript costs nothing. Every new account gets $50 in credit with no credit card, and sandbox keys never bill.
Hear or Omni for a French voice product?
Hear suits teams that already own the reasoning and the voice and just need fast French text. Omni suits teams that want PyAI to listen, detect the turn, reason and speak in one session, with a turn detector that holds while someone pauses mid thought or reads digits.
Prices checked November 2026. Hear figures from pyai.com/benchmarks.