Crate crispasr 

Safe Rust wrapper for CrispASR speech recognition.

Quick start

use crispasr::Session;

let sess = Session::open("model.gguf").unwrap();
let pcm = vec![0.0f32; 16000]; // 1s of silence
let segments = sess.transcribe(&pcm).unwrap();
for seg in &segments {
    println!("[{:.1}s - {:.1}s] {}", seg.start, seg.end, seg.text);
}

Structs

AlignedWord

ChatGenerateOptions

Per-call sampler options for the *_with_options generate entry points.

ChatMessage

One turn in a conversation.

ChatOptions

Per-session open options for ChatSession::open_with_options.

ChatSession

A chat / LLM session over a GGUF model.

CrispASRDeprecated

A loaded CrispASR model (whisper-only, legacy API).

CtcLogits

Per-frame CTC logits captured from a CTC backend.

DiarizeOptions

Construct via Default and set fields as needed — the struct grows alongside the append-only C ABI (#332).

DiarizeSegment

One ASR segment passed to diarize_segments. Caller fills t0 / t1 (seconds) from the upstream transcribe result; the diarizer writes the zero-based speaker index into speaker (-1 means the method had no info to pick).

DiarizeTurn

A speaker turn the diarizer derived from the AUDIO, independent of the caller’s segment grid (#395).

KokoroResolved

Result of kokoro_resolve_for_lang. Mirrors the Python wrapper’s KokoroResolved dataclass.

LidResult

Mic

Library-level microphone handle. The user-supplied callback is invoked from miniaudio’s audio thread with mono float32 PCM in [-1, 1]. Keep the callback short and non-blocking — for ASR, queue the audio and feed Stream::feed from another thread.

Parakeet

Direct Parakeet ASR context with word- and token-level timestamps.

ParakeetResult

Parakeet transcription result with word and token accessors.

PuncModel

BERT-based punctuation restoration model (FireRedPunc).

PyannoteCache

Pre-computed pyannote-seg posteriors over a full audio buffer.

RegistryArtifact

One file in a backend’s canonical default download bundle.

RegistryBundle

The exact artifact bundle downloaded by -m auto.

RegistryEntry

Known-model registry entry.

Segment

A transcription segment with timing information.

Session

SessionSegment

A segment of a unified-session transcription.

SessionWord

Word-level timing (populated by backends that produce it).

SpeakerDB

Speaker database wrapper.

SpeakerEmbedder

Pluggable speaker-embedding model. Wraps the crispasr_speaker_embedder_*_abi family in a safe Rust struct.

Stem

A loaded session over a CrispASR model of any backend. One stem from Session::separate: its name (vocals, drums, …) and its interleaved-stereo PCM at Session::separate_sample_rate.

Stream

Streaming-decoder handle returned by Session::stream_open. Feed PCM with Stream::feed, pull text with Stream::get_text, finalize with Stream::flush. Auto-closes on drop.

StreamingUpdate

One commit from a streaming session — the latest concatenated text plus its absolute audio-time bounds.

TextLidResult

Result of text_detect_language. Label format depends on the loaded GGUF: CLD3 returns ISO 639-1 ("en", "de", "zh-Latn") across 109 labels; GlotLID-V3 / LID-176 fastText return ISO 639-3 with a script tag ("eng_Latn", "sco_Latn") across 2102 or 176 labels respectively. Callers needing ISO 639-1 normalisation must do it on their side — the dispatcher preserves the model’s native space because the script tag carries real information (e.g. zh-Latn ≠ zh-Hans).

TranscribeOptions

Options for transcribe_pcm_with_options. Leave defaults for standard Whisper behaviour; set vad: true + vad_model_path for built-in Silero VAD, or tdrz: true with a .en.tdrz model for speaker-turn markers.

VadOptions

Tunables for Session::transcribe_vad. Defaults mirror crispasr’s whisper_vad_default_params plus the max-chunk fallback the shared library uses to bound encoder cost on long audio.

Enums

ChatError

Why a chat call failed.

DiarizeMethod

Mirrors the append-only CrispasrDiarizeMethod C enum, so new methods may appear in minor releases — match with a _ arm (#332).

LidMethod

Mirrors the append-only C LID-method enum, so new methods may appear in minor releases — match with a _ arm (#332).

RegistryArtifactKind

Role of one artifact in a canonical model download bundle.

Functions

agglomerative_cluster

Agglomerative single-linkage cosine clustering on L2-normalized speaker embeddings.

align_words

Run CTC / forced-aligner word timings for a transcript + audio pair.

cache_dir

Return the CrispASR cache directory (creating it if missing).

cache_ensure_file

Download filename from url into the CrispASR cache — or return the cached path if already present. Pass None for cache_dir_override to use the platform default.

detect_language_pcm

Run language identification on a 16 kHz mono float PCM buffer.

diarize_segments

Assign a speaker index to each of segs, mutating each DiarizeSegment::speaker in place.

diarize_segments_with_turns

Same as diarize_segments, and additionally returns the speaker turns the method derived from the audio (#395).

enhance_audio_rnnoise

RNNoise audio enhancement on 48 kHz mono PCM.

kokoro_lang_has_native_voice

Whether lang has a native Kokoro voice.

kokoro_lang_is_german

Whether lang is German (Kokoro phoneme selection).

kokoro_resolve_for_lang

Resolve the kokoro model + fallback voice for lang. Mirrors what the CLI does for --backend kokoro -l <lang> (PLAN #56 opt 2b).

lcs_dedup_prefix_count

Chunk-boundary LCS dedup: returns the number of leading tokens of curr_tokens to drop to remove overlap with prev_tail_tokens.

list_known_models

Look up the canonical GGUF for a backend (whisper, parakeet, canary, voxtral, voxtral4b, granite, granite-4.1, qwen3, cohere, wav2vec2). Returns None on miss. List every backend name in the registry, in declaration order.

mic_default_device_name

Human-readable name of the default capture device, or empty string if no input device is available.

registry_default_bundle

Return the backend’s exact canonical -m auto artifact bundle.

registry_lookup

registry_lookup_by_filename

Look up by filename (exact match, then fuzzy substring).

text_detect_language

Detect the language of a UTF-8 text string via the internal text_lid_dispatch (peek the GGUF’s general.architecture → route to CLD3 or fastText).

titanet_cosine_sim

TitaNet cosine similarity between two embeddings.

vad_segments

Run standalone VAD returning speech spans in centiseconds.

vad_slices

Run unified VAD dispatcher returning speech spans in seconds.