1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
//! # Voxtral Realtime
//!
//! Mistral's Voxtral Realtime speech recognizer
//! (`mistralai/Voxtral-Mini-4B-Realtime-2602`): a causal audio encoder whose
//! frames are added, one group per 80 ms, to the token embeddings of a
//! Mistral text model that writes the transcript a fixed delay behind the
//! audio. Every part is causal, so the model streams natively: each 80 ms of
//! audio is one decoder step.
//!
//! - [`blocks`]: the model, [`VoxtralRealtime`], built from a
//! [`VoxtralRealtimeConfig`]: a
//! [`VoxtralAudioEncoder`](blocks::VoxtralAudioEncoder) and a
//! [`VoxtralTextModel`](blocks::VoxtralTextModel) on the shared
//! [`CausalDecoderBlock`](crate::blocks::transformers::decoders::causal::CausalDecoderBlock),
//! and a [`VoxtralProjector`](blocks::VoxtralProjector) between them.
//! - [`VoxtralFrontEnd`]: the feature extractor, Whisper's log-mels under a
//! fixed level.
//! - [`VoxtralTokenLayout`]: the stream's special tokens, its delay (480 ms by
//! default; any multiple of 80 ms up to 1200, or 2400), the padding around an
//! utterance, and the prompt.
//! - [`VoxtralGreedyDecoder`]: greedy decoding through a KV cache, one id per
//! position, each position's audio added to its token
//! ([`VoxtralDecodeStep`]).
//! - [`VoxtralTranscript`]: a decode's ids as timed [`VoxtralWord`]s and text.
//! - [`VoxtralConstruct`]: a model with its front end, vocabulary
//! ([`TekkenVocabulary`](crate::kits::support::tokens::TekkenVocabulary)),
//! layout and decoder; [`transcribe`](VoxtralConstruct::transcribe) one
//! utterance.
//! - [`driver`]: [`VoxtralEngine`](driver::VoxtralEngine), a construct as the
//! [`asr_driver`](crate::kits::speech::asr::driver) core's engine, each
//! window decoded whole and laid out a segment per timed word; a
//! [`VoxtralStreamDriver`](driver::VoxtralStreamDriver) streams audio through
//! it, under any
//! [`EmissionPolicy`](crate::kits::speech::asr::driver::EmissionPolicy).
//! - [`pretrained`] (features `store` and `cache`): the named checkpoint,
//! `mistralai/voxtral-mini-4b-realtime-2602` the default, `hf:` repos, and
//! the factory that loads one into a construct.
//! - [`weights`] (feature `store`): loading tensors named as the published
//! `model.safetensors` names them, and reading its `config.json`.
pub use *;
pub use *;
pub use *;
pub use *;
pub use *;
pub use ;