LecSync vs Mistral Voxtral Realtime: Live Translation, Speakers and Price

Mistral now sells live transcription as Voxtral Mini Transcribe Realtime 3, a WebSocket model released in public preview on October 20, 2026. It streams text as people talk, lets you pick a delay between 80 ms and 2.4 seconds, and can label speakers or detect the end of a turn. It does not translate the stream, and it does not write minutes.

The LecSync Realtime Speech API is built for the other half of that job. One connection returns preview and final captions, a translation between any two of 60 languages, a speaker and audio-aligned timestamps on every final segment, and running minutes. It costs $0.006 per audio minute ($0.36 per hour) with all of that switched on.

The Mistral figures below come from its official docs, read on October 21, 2026: the API pricing page, the Realtime 3 model card, the realtime transcription guide, speaker diarization, supported languages and audio in chat completions.

The short answer

For same-language captions or a voice agent that needs turn detection, Voxtral is a reasonable choice. Realtime 3 costs $0.008 per minute, or $480 for 1,000 hours. The older voxtral-mini-transcribe-realtime-2602 model is still listed at $0.006 per minute, the same as LecSync, and its weights are published under Apache 2.0 if you would rather host it yourself.

Once captions have to be translated, carry speaker times, or update with better text after a first guess, you are building on top of Voxtral. LecSync does those things on the same socket for $360 per 1,000 hours, which is $120 less than Realtime 3 before any translation is added.

How each API works

LecSync. Your server calls POST https://api.lecsync.com/v1/realtime/connect with the audio format, a language pair and the features you want, then streams 16 kHz mono PCM to the URL it gets back. The socket sends preview text while someone is still talking, then a final segment once the sentence settles. Translations arrive paired with their segment ID. With speakers: true, each final carries a speaker label and startMs/endMs on the recording's own clock. With digest: true, minutes arrive section by section. Language is detected automatically and the translation direction is picked per segment, so a Chinese and English meeting runs on one stream. The first word shows up about 0.28 seconds after speech starts, and translation adds roughly 88 ms.

Mistral Voxtral Realtime 3. You open wss://api.mistral.ai/v1/audio/transcriptions/realtime?model=voxtral-mini-transcribe-realtime-3 with your API key, wait for session.created, and send one session.update. That message sets the PCM sample rate (16 kHz by default), target_streaming_delay_ms, an optional language code, context_bias terms, and either diarize or turn_detection. Audio then goes up as base64 chunks through input_audio.append, and text comes back as transcription.text.delta events until transcription.done returns the full transcript.

Three details in Mistral's guide matter for caption products:

• Every delta is final. The docs say a delta "never corrects earlier text", so your UI only appends. To show fast text first and replace it with more accurate text later, Mistral tells you to run two sessions on the same audio, which means paying for the minutes twice.
• Settings lock when audio starts. A session.update after the first chunk returns an error. Changing the language or delay means opening a new session.
• Speakers or turns, not both. diarize and turn_detection cannot share a session.

Speaker labels in a live session

Voxtral Realtime 3 added live diarization. With diarize: true, the session sends transcription.segment events instead of deltas. Each event holds a few words and a speaker_id, but Mistral's own example shows start and end as null. You can show who is talking, but you cannot jump to that line in the recording from the live event.

You can also pass up to 8 reference clips of 3 to 15 seconds each, so known voices come back by name. The total number of speakers is capped at 10. The docs also state that overlapping speech is not supported in this version.

LecSync attaches a speaker and millisecond timestamps to every final segment. Those times stay on one recording clock even when the upstream connection reconnects, so they can be used for playback as they are. The details are in building a meeting client with speakers and timestamps.

Translation is a separate call

Mistral's language page lists translation as a task of audio in chat completions, not of the realtime model. You send a recorded file, base64-encoded, to a chat model such as voxtral-small-latest with an instruction like "Translate this call into French", and get text back. The guide notes that the answer carries no timestamps, so a translated line cannot be matched to a moment in the audio.

For live translated captions on Voxtral, the usual design is to collect each finished sentence from the realtime stream and send it to a text model. You then own sentence boundaries, retries and pairing each translation with its source line. How much of that work adds up is covered in built-in translation vs a DIY STT + MT pipeline.

Side-by-side facts (checked October 21, 2026)

The 2602 column is filled in only where Mistral's docs describe that model. Its card calls it "optimized for live transcription purposes only", and Mistral's migration guide lists realtime diarization and turn detection as new in version 3.

What 1,000 hours cost

The prices are list rates per audio minute: 60,000 minutes times $0.006 or $0.008. The translation rows assume about 12,000 input tokens and 15,000 output tokens per hour of English speech, sent to Mistral Small 4 at $0.15 and $0.60 per million tokens. Those token counts are estimates. They show that model cost is small here; the bigger cost is the pipeline you have to build and run. For more vendors on the same basis, see what 1,000 hours cost once translation, speakers and summaries are added.

When Voxtral is the better pick

• Voice agents. Built-in turn events (turn.start, turn.end and an eager end) help an agent decide when to answer. Turn detection is not part of LecSync's API.
• Tuning delay against accuracy. You can go as low as 80 ms or as high as 2.4 seconds per session.
• Naming known speakers. Reference clips let a returning voice come back with its name instead of speaker_1.
• Self-hosting. The 2602 realtime model ships open weights under Apache 2.0.
• Staying inside Mistral. If your chat, OCR and TTS already run on Mistral, one bill and one key can be worth a lot.

When LecSync fits better

• Translated live captions. Any pair of 60 languages, on the same socket, with each translation tied to its segment.
• Two-way meetings. One stream translates Chinese to English and English to Chinese, deciding the direction per segment.
• Captions that improve in place. Preview text appears immediately and settles into a final, without a second paid session.
• Speaker times you can play back. Every final has a speaker and startMs/endMs on the recording clock.
• Long, unstable sessions. LecSync caps server-side backlog at 2 seconds, so captions recover within a second after a slowdown, as described in how LecSync caps backlog at 2 seconds.
• Minutes while people are still talking. digest: true pushes confirmed sections during the session. See live digest vs post-call summarization.

If you are also looking at other speech vendors, see LecSync vs ElevenLabs Scribe Realtime and LecSync vs AssemblyAI and OpenAI Realtime.

FAQ

How much does Mistral Voxtral realtime transcription cost?

As of October 21, 2026, Mistral lists Voxtral Mini Transcribe Realtime 3 at $0.008 per audio minute, or $0.48 per hour. The previous realtime model, voxtral-mini-transcribe-realtime-2602, is listed at $0.006 per minute. Neither price includes translation. LecSync charges $0.006 per minute with translation, speakers and live minutes included.

Does Voxtral Realtime translate speech live?

No. Mistral's docs treat translation as a task of audio in chat completions: you send a recorded file to a chat model and get translated text back, without timestamps. For live translated captions, you would pass each finished sentence from the realtime stream to a text model yourself. LecSync translates on the same socket.

Does Voxtral Realtime support speaker diarization?

Yes, since Realtime 3. Set diarize: true and the session returns transcription.segment events with a speaker_id, up to 10 speakers. In live sessions, start and end are null, overlapping speech is not supported, and diarization cannot be combined with turn detection. LecSync labels every final segment with a speaker and timestamps.

Is Voxtral Realtime cheaper than LecSync?

For same-language captions on the 2602 model, the price is the same: $0.006 per minute. Realtime 3 costs a third more at $0.008. If you self-host the open 2602 weights, your cost depends on your own GPUs. Once you add translation, live minutes or a second session for corrected text, LecSync's single rate comes out lower.

Bottom line

Pick Voxtral Realtime 3 for a voice agent that needs turn detection, or when you want to tune the delay yourself and already build on Mistral. Pick the 2602 model if open weights matter. Pick LecSync when captions have to be translated, labelled by speaker with times you can play back, corrected in place, and summarized during the meeting, all at $0.006 per minute. You can open a test stream from the developers page with one POST /v1/realtime/connect.