Skip to main content
ClovisSTT/realtime

ClovisSTT/realtime

Mistral — Voxtral Mini 4B Realtime 2602Audio transcription

Turn streaming audio into text as the speaker talks, for dictation, live captions or voice input for an assistant.

Capabilities

InputWhat you can send
InputMono PCM16 audio at 16 kHz, Base64-encoded
OutputWhat the model produces
OutputIncremental text followed by a final transcript
TransportSecure WebSocket (WSS)
Model languagesFrenchEnglishGermanSpanishItalianPortugueseDutchRussianArabicHindiChineseJapaneseKorean

Compatible endpoints

Realtime — WebSocket→

The model behind ClovisSTT/realtime​

ClovisSTT/realtime is the name to use in the Clovis API. It corresponds to Voxtral Mini 4B Realtime 2602, Mistral AI’s streaming speech recognition model. It is described as a 4-billion-parameter-class model and released under the Apache 2.0 license. Its architecture processes audio as it arrives.

Use cases​

  • Dictation: turn speech into text in an application.
  • Live captions: display a transcript during a conversation.
  • Meetings and interviews: collect a transcript for later processing.
  • Voice assistants: send recognized text to a conversational model.

ClovisSTT produces text from speech. To respond to a user or summarize a transcript, your application can then call ClovisChat or ClovisLLM.

Languages and latency​

The 13 model languages​

ClovisSTT/realtime can recognize and transcribe speech in the following 13 languages:

French · frEnglish · enGerman · deSpanish · esItalian · itPortuguese · ptDutch · nlRussian · ruArabic · arHindi · hiChinese · zhJapanese · jaKorean · ko

The codes identify languages here; they are not parameters to add to the WebSocket events in this guide. The transcript stays in the spoken language: transcription does not translate it. Accuracy can vary with the language, accent and recording quality.

Transcription delay​

Mistral recommends a transcription delay of 480 ms as a balance between responsiveness and accuracy. This model delay is not an end-to-end latency guarantee for the Clovis API: the network, audio chunk size and service load also affect latency.

The following recommendations apply to deploying Voxtral with vLLM. Server parameters are managed by Clovis; they are not fields to add to client WebSocket events.

  • Temperature: use 0.0 for transcription.
  • Maximum length: size --max-model-len for the recording duration. One text token corresponds to roughly 80 ms of audio. For a one-hour meeting, allow at least 45,000 tokens: 3,600 / 0.08 = 45,000. Mistral recommends the default 131,072 tokens, corresponding to about 3 hours. Streaming can theoretically continue indefinitely, but server memory allocations, including RoPE positions, impose practical limits. These values do not guarantee session duration on the Clovis API.
  • Transport: prefer WebSockets to stream audio and receive transcripts continuously.
  • Transcription delay: prefer 480 ms. For a custom deployment, the transcription_delay_ms field in tekken.json can be set to a multiple of 80 ms between 80 and 1,200 ms, or to 2,400 ms. This is a server-side setting.

Transcription quality​

Accuracy depends on the recording: noise, overlapping voices, proper nouns and specialized vocabulary can affect the result. Review transcripts before using them in meeting minutes.

The shared test covers three English clips from hf-internal-testing/librispeech_asr_dummy: two pass the comparison threshold, while the third fails the recall threshold (67% against an 80% threshold), with reported differences in capitalization and punctuation. This check validates integration on these clips; it does not measure the model’s overall accuracy. To evaluate your use case, test representative recordings and normalize capitalization and punctuation if they are not part of your criteria.