transcribe-cpp-sys 0.0.8

Native FFI bindings for transcribe.cpp, a C/C++ speech-to-text library built on ggml
Documentation 
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
69
70
71
72
73
74

// arch/parakeet/capabilities.cpp - Parakeet capability defaults.
//
// Per the 2B revision after the variant-dispatch review (see
// RESUME.md "Decisions still load-bearing" -> "stt.variant is
// descriptive metadata, not a behavioral discriminator"):
//
//   - The previous variant-driven resolver is gone. Capability
//     resolution is now KV-driven via transcribe::read_capability_kv.
//   - This file holds only the family-default population, applied
//     before the KV reader runs. KV present overrides the family
//     default; KV absent leaves the family default in place. PLAN.md
//     "GGUF KV is authoritative; if a key is absent, the architecture
//     supplies a default" — this function supplies those defaults.
//
// What stays here vs what goes to KV:
//
//   - native_sample_rate is set here, not exposed via KV. Parakeet's
//     feature extractor is a fixed 16 kHz mel bank for every
//     published variant; treating it as a KV-overridable field would
//     just create a footgun. If a future variant truly used a
//     different rate, the right answer would be to add a separate
//     parakeet handler subtree, not to flip a KV.
//
//   - supports_translate defaults to false here. Strictly speaking
//     the KV reader could overwrite it, but no published Parakeet
//     variant has a translate task head and the converter is not
//     expected to emit stt.capability.translate for this family. The
//     default sticks because nothing overrides it.
//
//   - supports_language_detect, supports_streaming come from KV. The
//     family default is false; v3 multilingual GGUFs are expected to
//     emit stt.capability.lang_detect = true and the KV reader will
//     pick it up.
//
//   - n_languages / languages come from general.languages via
//     read_languages_kv (called by the family handler after this
//     function runs). Absent leaves the model in the documented
//     "information gap, not a claim" state.

#include "parakeet.h"

namespace transcribe::parakeet {

void apply_family_invariants(transcribe_model & model) {
    transcribe_capabilities & caps = model.caps;

    // Parakeet's feature extractor is a fixed 16 kHz mel bank (NeMo
    // AudioToMelSpectrogramPreprocessor). Every published variant — v2
    // English and v3 multilingual — uses it. The CLI rejects non-
    // 16 kHz WAVs at load time so this number is the contract.
    caps.native_sample_rate = 16000;

    // Family default. No published Parakeet variant has a translate
    // task head; NeMo ships translation via Canary, not Parakeet.
    // KV is authoritative per PLAN.md, so a converter that explicitly
    // wrote stt.capability.translate could in principle override
    // this — but for Parakeet we never expect that to happen.
    caps.supports_translate = false;

    // Parakeet's TDT decoder emits one token per acoustic frame the
    // encoder kept, with step_at_emit carrying the frame index.
    // run() converts that to per-token t0/t1 and aggregates up to
    // word- and segment-level timings. TOKEN is the finest the
    // family can honestly produce, and that is what we advertise.
    caps.max_timestamp_kind = TRANSCRIBE_TIMESTAMPS_TOKEN;

    // Abort callback is honored at the top of each run. Other
    // long-form / temperature-fallback / initial-prompt features
    // belong to autoregressive families and don't apply here. No
    // runtime PNC/ITN control.
    transcribe::set_feature(&model, TRANSCRIBE_FEATURE_CANCELLATION, true);
}

} // namespace transcribe::parakeet