Crate crispasr_sys 

Raw FFI bindings to CrispASR. Mirrors the public C API in include/whisper.h.

Structs

CrispasrAlignResult

Opaque result handle for crispasr_align_words_abi. Must be freed with crispasr_align_result_free.

CrispasrChatError

Out-parameter for every chat entry point that can fail (may be null). Left untouched on success; on failure code is non-zero and message holds a NUL-terminated diagnostic.

CrispasrChatGenerateParams

Per-call, sampler-level generate params. Fill via crispasr_chat_generate_params_default before overriding fields.

CrispasrChatMessage

One turn of a conversation. role is “system”, “user”, “assistant” or “tool”; both pointers must stay valid for the duration of the call.

CrispasrChatOpenParams

Per-session, model-level open params. Fill via crispasr_chat_open_params_default before overriding fields — the C side reads every one of them.

CrispasrChatSession

Opaque handle returned by crispasr_chat_open. Free with crispasr_chat_close.

CrispasrDiarizeOptsAbi

ABI options for crispasr_diarize_segments_abi. method is a value in 0..4: 0 = Energy, 1 = Xcorr, 2 = VadTurns, 3 = Pyannote, 4 = FoxNose. pyannote_model_path is required for Pyannote, foxnose_embedder_path for FoxNose; each is ignored otherwise.

CrispasrDiarizeSegAbi

ABI segment for crispasr_diarize_segments_abi. Caller fills t0_cs / t1_cs; the diarizer writes speaker (-1 if unassigned).

CrispasrMic

Opaque microphone handle returned by crispasr_mic_open. Must be freed with crispasr_mic_close. (PLAN #62d)

CrispasrSession

Opaque handle returned by crispasr_session_open.

CrispasrSessionResult

Opaque result handle returned by crispasr_session_transcribe. Must be freed with crispasr_session_result_free.

CrispasrStream

Opaque streaming-decoder handle returned by crispasr_session_stream_open. Must be freed with crispasr_stream_close. (PLAN #62)

CrispasrVadAbiOpts

Tunables for crispasr_session_transcribe_vad. Mirrors crispasr’s whisper_vad_params plus the max-chunk fallback used to bound encoder cost on long audio. Pass a null pointer to use defaults.

WhisperContext

Opaque context handle.

WhisperContextParams

Opaque context params handle.

WhisperFullParams

Opaque params handle (allocated by whisper_full_default_params_by_ref).

WhisperState

Opaque state handle.

Constants

CRISPASR_CHAT_ERR_ABORTED

The one error code on the chat ABI with a stable, documented meaning: a registered abort callback stopped the run. Every other non-zero CrispasrChatError::code is a diagnostic aid — read message.

CRISPASR_SAMPLING_BEAM_SEARCH

CRISPASR_SAMPLING_GREEDY

Sampling strategy.

Functions

crispasr_align_result_free^⚠

crispasr_align_result_n_words^⚠

crispasr_align_result_word_t0^⚠

crispasr_align_result_word_t1^⚠

crispasr_align_result_word_text^⚠

crispasr_align_words_abi^⚠

Shared CTC / forced-aligner word timings (0.4.7+). Pass any aligner_model path — filenames containing “forced-aligner” / “qwen3-fa” / “qwen3-forced” go through the Qwen3-ForcedAligner path; everything else uses canary-ctc. Returns a handle the caller must free with crispasr_align_result_free. Returns null on failure.

crispasr_c_api_version^⚠

crispasr_cache_dir_abi^⚠

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

crispasr_cache_ensure_file_abi^⚠

Shared HF download + cache (0.4.8+). Writes the resolved path into out_buf. Returns 0 on success, -1 on invalid args, 1 on download failure, 2 when the output buffer is too small.

crispasr_chat_ai_disclosure_text^⚠

Canonical EU AI Act Art. 50(1) “you are talking to an AI” wording. Static string — never null, never freed.

crispasr_chat_close^⚠

crispasr_chat_count_tokens^⚠

Prompt tokens a FRESH session prefills for messages — chat template, BOS, and the trailing generation prompt included. Negative on failure, with err filled.

crispasr_chat_generate^⚠

Returns a malloc’d UTF-8 string (free with crispasr_chat_string_free) or null on failure / abort.

crispasr_chat_generate_params_default^⚠

crispasr_chat_generate_stream^⚠

0 on clean completion (including stop-sequence / EOG termination), CRISPASR_CHAT_ERR_ABORTED when the abort callback stopped it, other non-zero on failure.

crispasr_chat_memory_estimate^⚠

Approximate working set in bytes for a GGUF chat model on disk, or 0 when it could not be estimated (err filled).

crispasr_chat_n_ctx^⚠

crispasr_chat_open^⚠

crispasr_chat_open_params_default^⚠

crispasr_chat_reset^⚠

crispasr_chat_set_abort_callback^⚠

Register cb on the session (null clears it). Takes the session lock, so calling it during a generation blocks rather than cancelling — register before starting one.

crispasr_chat_string_free^⚠

crispasr_chat_template_name^⚠

crispasr_detect_backend_from_gguf^⚠

crispasr_detect_language^⚠

crispasr_detect_language_pcm^⚠

Shared language identification (0.4.6+). method is 0 for whisper, 1 for silero. model_path is required. Fills out_lang_buf with a null-terminated ISO 639-1 code. Returns 0 on success, -1 on invalid args, 1 on model / detect failure, 2 when the output buffer is too small.

crispasr_diarize_segments_abi^⚠

Shared speaker diarization (0.4.5+). Writes a zero-based speaker index into each segs[i].speaker. Returns 0 on success, 1 on Pyannote model load failure, -1 on invalid args.

crispasr_enhance_audio_rnnoise^⚠

crispasr_kokoro_lang_has_native_voice_abi^⚠

crispasr_kokoro_lang_is_german_abi^⚠

crispasr_kokoro_resolve_fallback_voice_abi^⚠

crispasr_kokoro_resolve_model_for_lang_abi^⚠

crispasr_lcs_dedup_prefix_count^⚠

crispasr_mic_close^⚠

crispasr_mic_default_device_name^⚠

crispasr_mic_open^⚠

crispasr_mic_start^⚠

crispasr_mic_stop^⚠

crispasr_parakeet_free^⚠

crispasr_parakeet_init^⚠

crispasr_parakeet_result_free^⚠

crispasr_parakeet_result_n_tokens^⚠

crispasr_parakeet_result_n_words^⚠

crispasr_parakeet_result_text^⚠

crispasr_parakeet_result_token_p^⚠

crispasr_parakeet_result_token_t0^⚠

crispasr_parakeet_result_token_t1^⚠

crispasr_parakeet_result_token_text^⚠

crispasr_parakeet_result_word_t0^⚠

crispasr_parakeet_result_word_t1^⚠

crispasr_parakeet_result_word_text^⚠

crispasr_parakeet_transcribe^⚠

crispasr_params_set_alt_n^⚠

crispasr_params_set_best_of^⚠

crispasr_params_set_detect_language^⚠

crispasr_params_set_initial_prompt^⚠

crispasr_params_set_language^⚠

crispasr_params_set_max_len^⚠

crispasr_params_set_max_tokens^⚠

crispasr_params_set_n_threads^⚠

crispasr_params_set_no_context^⚠

crispasr_params_set_print_progress^⚠

crispasr_params_set_print_realtime^⚠

crispasr_params_set_print_special^⚠

crispasr_params_set_print_timestamps^⚠

crispasr_params_set_single_segment^⚠

crispasr_params_set_split_on_word^⚠

crispasr_params_set_suppress_blank^⚠

crispasr_params_set_tdrz^⚠

crispasr_params_set_temperature^⚠

crispasr_params_set_token_timestamps^⚠

crispasr_params_set_translate^⚠

crispasr_params_set_vad^⚠

crispasr_params_set_vad_min_silence_ms^⚠

crispasr_params_set_vad_min_speech_ms^⚠

crispasr_params_set_vad_model_path^⚠

crispasr_params_set_vad_threshold^⚠

crispasr_pcm_free^⚠

crispasr_punc_free^⚠

crispasr_punc_free_text^⚠

crispasr_punc_init^⚠

crispasr_punc_process^⚠

crispasr_pyannote_cache_apply_abi^⚠

Score segs against the cached posteriors. slice_t0_cs is the absolute centisecond at which the cache buffer starts (typically 0 — the cache covers the whole input audio).

crispasr_pyannote_cache_compute_abi^⚠

Pre-compute pyannote-seg posteriors over a full audio buffer. Returns an opaque cache or null on failure. Free with crispasr_pyannote_cache_free_abi.

crispasr_pyannote_cache_free_abi^⚠

crispasr_registry_default_bundle_artifact_abi^⚠

Read one default-bundle artifact by index. 0 = success.

crispasr_registry_default_bundle_info_abi^⚠

Describe the exact canonical artifact bundle downloaded by -m auto. Returns its artifact count, 0 on miss, or a negative argument/buffer error.

crispasr_registry_list_backends_abi^⚠

Shared known-model registry lookup by filename (exact then fuzzy).

crispasr_registry_lookup_abi^⚠

Shared known-model registry lookup by backend. 0 = hit, 1 = miss.

crispasr_registry_lookup_by_filename_abi^⚠

crispasr_session_accept_marking_responsibility^⚠

crispasr_session_available_backends^⚠

Write a comma-separated list of backend names the loaded dylib was built with. Returns the number of bytes written (not counting NUL) or a negative error.

crispasr_session_backend^⚠

crispasr_session_close^⚠

crispasr_session_detect_language^⚠

crispasr_session_detected_language^⚠

crispasr_session_get_speaker_name^⚠

crispasr_session_input_channels^⚠

crispasr_session_input_sample_rate^⚠

crispasr_session_is_custom_voice^⚠

crispasr_session_is_voice_design^⚠

crispasr_session_kokoro_clear_phoneme_cache^⚠

crispasr_session_n_speakers^⚠

crispasr_session_n_vocab^⚠

crispasr_session_open^⚠

crispasr_session_open_explicit^⚠

crispasr_session_open_with_params^⚠

crispasr_session_output_channels^⚠

crispasr_session_output_sample_rate^⚠

crispasr_session_result_free^⚠

crispasr_session_result_logits^⚠

crispasr_session_result_n_logit_frames^⚠

crispasr_session_result_n_logit_vocab^⚠

crispasr_session_result_n_segments^⚠

crispasr_session_result_n_words^⚠

crispasr_session_result_segment_no_speech_prob^⚠

crispasr_session_result_segment_t0^⚠

crispasr_session_result_segment_t1^⚠

crispasr_session_result_segment_text^⚠

crispasr_session_result_word_alt_p^⚠

crispasr_session_result_word_alt_text^⚠

crispasr_session_result_word_n_alts^⚠

crispasr_session_result_word_p^⚠

crispasr_session_result_word_t0^⚠

crispasr_session_result_word_t1^⚠

crispasr_session_result_word_text^⚠

crispasr_session_separate^⚠

crispasr_session_separate_n_stems^⚠

crispasr_session_separate_sample_rate^⚠

crispasr_session_separate_stem^⚠

crispasr_session_separate_stem_name^⚠

crispasr_session_set_alt_n^⚠

crispasr_session_set_ask^⚠

crispasr_session_set_beam_size^⚠

crispasr_session_set_best_of^⚠

crispasr_session_set_cfg_weight^⚠

crispasr_session_set_codec_path^⚠

crispasr_session_set_do_sample^⚠

crispasr_session_set_exaggeration^⚠

crispasr_session_set_fallback_thresholds^⚠

crispasr_session_set_frequency_penalty^⚠

crispasr_session_set_g2p_dict^⚠

crispasr_session_set_grammar_text^⚠

crispasr_session_set_hotwords^⚠

crispasr_session_set_instruct^⚠

crispasr_session_set_length_scale^⚠

crispasr_session_set_max_new_tokens^⚠

crispasr_session_set_max_speech_tokens^⚠

crispasr_session_set_min_p^⚠

crispasr_session_set_min_speech_tokens^⚠

crispasr_session_set_progress_callback^⚠

0.10.3+ (issue #208): register a per-session progress callback for long-form (chunked) transcription. Fired once per finished window with (processed_samples, total_samples, user_data); processed is monotonic and reaches total on the last window. Invoked on the transcribe thread. Pass None/null cb to clear.

crispasr_session_set_punc_model^⚠

crispasr_session_set_punctuation^⚠

crispasr_session_set_repetition_penalty^⚠

crispasr_session_set_return_logits^⚠

crispasr_session_set_sensitivity^⚠

crispasr_session_set_source_language^⚠

crispasr_session_set_speaker_id^⚠

crispasr_session_set_speaker_identity^⚠

crispasr_session_set_speaker_name^⚠

crispasr_session_set_target_language^⚠

crispasr_session_set_temperature^⚠

crispasr_session_set_top_k^⚠

crispasr_session_set_top_p^⚠

crispasr_session_set_translate^⚠

crispasr_session_set_tts_noise_temp^⚠

crispasr_session_set_tts_num_candidates^⚠

crispasr_session_set_tts_phonemes^⚠

crispasr_session_set_tts_reference_language^⚠

crispasr_session_set_tts_seed^⚠

crispasr_session_set_tts_steps^⚠

crispasr_session_set_voice^⚠

crispasr_session_set_whisper_decode_extras^⚠

crispasr_session_speech_to_speech^⚠

crispasr_session_stream_open^⚠

crispasr_session_synthesize^⚠

crispasr_session_synthesize_raw^⚠

crispasr_session_token_text^⚠

crispasr_session_transcribe^⚠

crispasr_session_transcribe_chunked^⚠

crispasr_session_transcribe_chunked_lang^⚠

0.8.7+: chunked-encode transcribe (issue #208). Forces the Parakeet backend through its bounded long-form path (overlapping short-window transcribe-and-merge for non-JA models, streamed encoder for the JA-only model) regardless of audio length, so long files transcribe in bounded time AND recover the sections a single full-length pass drops. chunk_seconds <= 0 keeps the per-model defaults; otherwise it sets the non-JA window length / the JA streamed window. overlap_seconds < 0 uses the default. For non-Parakeet backends the chunk params are inert and this matches crispasr_session_transcribe_lang.

crispasr_session_transcribe_lang^⚠

0.4.9+: language-aware session transcribe. language is an ISO 639-1 code or null/empty to keep the backend’s historical default. Backends that accept a source-language hint (whisper, canary, cohere, voxtral, voxtral4b) honour it; others ignore silently.

crispasr_session_transcribe_vad^⚠

VAD-driven session transcribe. Runs Silero VAD on the PCM buffer, merges short / overlong speech slices, stitches them into one contiguous buffer with 0.1s silence gaps, calls the backend once, then remaps segment + word timestamps back to original-audio positions.

crispasr_session_transcribe_vad_lang^⚠

0.4.9+: language-aware VAD transcribe (same semantics as the language kwarg on crispasr_session_transcribe_lang).

crispasr_session_translate_text^⚠

crispasr_session_translate_text_free^⚠

crispasr_speaker_cluster_abi^⚠

Agglomerative single-linkage cosine clustering. embeddings is a row-major n × dim buffer of (ideally L2-normalized) vectors. labels_out receives one cluster ID per input in [0, k). Returns the cluster count k, or -1 on invalid arguments.

crispasr_speaker_db_count^⚠

crispasr_speaker_db_enroll^⚠

crispasr_speaker_db_free^⚠

crispasr_speaker_db_load^⚠

crispasr_speaker_db_match^⚠

crispasr_speaker_embedder_dim_abi^⚠

Output embedding dimension (e.g. 192 for TitaNet, 512 for IndexTTS-BigVGAN).

crispasr_speaker_embedder_embed_abi^⚠

Extract one embedding. out must hold at least dim() floats. Returns 1 on success, 0 if the model rejected the input.

crispasr_speaker_embedder_free_abi^⚠

crispasr_speaker_embedder_make_abi^⚠

Build a pluggable speaker embedder. model_spec is one of "auto", "titanet", "indextts", "indextts-bigvgan", "ecapa", or a .gguf path. Returns null on failure.

crispasr_speaker_embedder_name_abi^⚠

crispasr_stream_close^⚠

crispasr_stream_feed^⚠

crispasr_stream_flush^⚠

crispasr_stream_get_text^⚠

crispasr_stream_open^⚠

crispasr_stream_set_live_decode^⚠

Toggle voxtral4b live-captions decode-during-feed (PLAN #7 phase 3). No-op for backends that don’t have audio-injection prompt decode. Set BEFORE the first feed for clean semantics.

crispasr_text_detect_language^⚠

crispasr_titanet_cosine_sim^⚠

crispasr_titanet_embed^⚠

crispasr_titanet_free^⚠

crispasr_titanet_init^⚠

crispasr_token_alt_id^⚠

crispasr_token_alt_p^⚠

crispasr_token_alt_text^⚠

crispasr_token_n_alts^⚠

crispasr_token_p^⚠

crispasr_token_t0^⚠

crispasr_token_t1^⚠

crispasr_vad_free^⚠

crispasr_vad_segments^⚠

crispasr_vad_slices^⚠

crispasr_watermark_detect^⚠

crispasr_watermark_embed^⚠

whisper_context_default_params_by_ref^⚠

whisper_free^⚠

whisper_free_context_params^⚠

whisper_free_params^⚠

whisper_full^⚠

whisper_full_default_params_by_ref^⚠

whisper_full_get_segment_no_speech_prob^⚠

whisper_full_get_segment_t0^⚠

whisper_full_get_segment_t1^⚠

whisper_full_get_segment_text^⚠

whisper_full_lang_id^⚠

whisper_full_n_segments^⚠

whisper_init_from_file_with_params^⚠

whisper_lang_id^⚠

whisper_lang_str^⚠

Type Aliases

CrispasrChatAbortCallback

Abort hook. Returns true to continue, false to abort — the whisper_encoder_begin_callback convention on the ASR surface, and the opposite of ggml’s own. Called on the generating thread before each prompt batch and each sampled token, and (CPU backend only) from inside a running compute graph, so it must be cheap and non-blocking. It must not re-enter the session that registered it — the session mutex is held.

CrispasrChatOnToken

Fired once per detokenised UTF-8 chunk during a streaming generate. The chunk pointer is valid only for the duration of the call. Option<...> so a null pointer clears the callback (C NULL).

CrispasrProgressCallback

Progress callback for long-form (chunked) transcription (issue #208). Option<...> so a null pointer clears the callback (C NULL).