Documented omissions: VAD, grammar, callbacks, DTW, low-level encode/decode, etc.

[uqio]

This umbrella issue catalogues whisper.cpp surface that is intentionally not wrapped by the safe whispercpp API. The omission is a deliberate scope choice — whispery (the in-house consumer) doesn't need these, and adding them speculatively expands the audited unsafe surface.

If you have a concrete use case for any item below, comment with your scenario. The decision isn't permanent — it's just deferred until a caller materialises.

See whispercpp/TODO.md § 1 for the full rationale.

Built-in VAD

Whisper.cpp ships its own Silero ONNX VAD. Whispery uses the silero crate upstream, which is the canonical path; re-wrapping whisper.cpp's VAD duplicates state.

Symbols: whisper_vad_*, whisper_full_params::vad, vad_model_path, vad_params, set_min_speech_duration_ms, set_max_speech_duration_s, set_min_silence_duration_ms, set_speech_pad_ms, set_threshold.

Grammar-constrained decoding

Pulls a sizeable struct hierarchy (whisper_grammar_element, rules, stacks) and a non-trivial ownership model. No caller has asked for it.

Symbols: whisper_full_params::grammar_rules, grammar_n_rules, grammar_i_start_rule, grammar_penalty, set_grammar, set_grammar_penalty, set_start_rule, whisper_grammar_*.

Translate task (audio → English)

Params::set_translate(true) is wrapped (1-line passthrough), but the full translate-task flow (token-id remapping, prompt seeding) is not exercised by any caller and we don't ship test coverage.

Tinydiarize input controls

Segment::speaker_turn_next() IS wrapped (1-byte read). Configuring --tdrz on the input side (set_tdrz_enable) is not — it requires a TDRZ-enabled checkpoint we don't ship, and whispery's diarization runs upstream via clustering on word ranges.

Symbols: whisper_full_params::tdrz_enable, set_tdrz_enable.

Lower-level encode/decode entry points

We expose state.full() only. The lower-level encode/decode flow (running the encoder, then decode token-by-token with custom sampling) is meaningful for research / custom samplers but doesn't fit whispery's pump architecture.

Symbols: whisper_encode, whisper_encode_with_state, whisper_decode, whisper_decode_with_state, whisper_get_logits, whisper_get_logits_from_state, whisper_set_mel, whisper_set_mel_with_state, whisper_pcm_to_mel, whisper_pcm_to_mel_with_state.

Mid-decode callbacks (progress / new segment / logits filter / encoder start)

Each requires the same trampoline discipline as the abort callback and adds another Box<dyn FnMut> field to Params. None is wired into whispery's chunk-at-a-time pump.

Symbols: set_progress_callback*, set_new_segment_callback*, set_filter_logits_callback*, set_start_encoder_callback*.

Related: see #3 for an iterator-based output API and #4 for an async variant — but for true streaming partial-result emission, the new-segment callback is the underlying mechanism and would need wrapping.

Global logging hooks

Whispery routes diagnostics through its own eprintln! / tracing layer. whisper_set_log_callback is a global hook that fires across all instances; mixing it with Rust logging frameworks needs more design than a 1:1 port.

Symbols: whisper_set_log_callback, set_debug_mode, whisper_log_callback.

Buffer-load and custom-loader constructors

We support Context::new(path, params) only. Loading from an in-memory buffer or via a custom whisper_model_loader is rare and adds lifetime/ownership complexity.

Symbols: whisper_init_from_buffer_with_params, whisper_init_with_params (custom loader), whisper_model_loader.

Direct beam_search / greedy field accessors

Use Params::new(SamplingStrategy::BeamSearch { beam_size, patience }) / Greedy { best_of }. Direct whisper_full_params::beam_search.beam_size / patience field access isn't exposed.

---

Removed from this list

• DTW token timestamps — landed on feat/dtw. Enable via ContextParams::with_dtw_token_timestamps(true) + with_dtw_aheads_preset(...); read per-token timing as Token::t_dtw() -> Option<i64>. See the README's DTW section for the full API.

[uqio]

DTW token timestamps removed from the omissions list — landed on feat/dtw (PR pending). Safe API:

• ContextParams::with_dtw_token_timestamps(true).with_dtw_aheads_preset(LargeV3Turbo) to enable
• Token::t_dtw() -> Option<i64> to read per-token timing (None = DTW unavailable for this token)
• required_dtw_mem_size_for(preset) to pre-compute the per-preset scratch budget
• All native abort paths inside the DTW helper converted to WhisperError::StateLost via the existing exception shim.

The other items remain deferred — comment on this issue if you have a concrete use case for any of them.