fix(stt): thread confidence thresholds into faster-whisper's own gate (#74178)

[PRATHAMESH75]

What & why

build_local_transcribe_kwargs() reads stt.local.no_speech_prob_threshold and
stt.local.logprob_threshold from config, but only uses them in Hermes' own
post-filter (_is_hallucinated_segment via _join_confident_segments). It never
passed them to faster-whisper's model.transcribe(), so the library's internal
defaults (no_speech_threshold=0.6, log_prob_threshold=-1.0) always applied and
silently dropped low-confidence segments before they ever reached our
post-filter. That made those two config knobs dead for the first (and decisive) gate.

Non-English speech decodes at a lower avg_logprob, so the English-tuned defaults
discard whole utterances — the caller gets an empty transcript even though the mic
captured audio and language detection succeeded (lang=zh, prob=1.000).

Fix

Map the same config values through to model.transcribe() so faster-whisper's
internal gate and Hermes' post-filter use identical thresholds. Reuses the existing
_confidence_thresholds() helper (same parsing + fallback the post-filter already
uses), so there's a single source of truth. Defaults are unchanged (0.6 / -1.0),
so behavior is byte-for-byte identical unless a user tunes them — this only makes
the previously-dead knobs actually reach the model.

Users hitting this on non-English speech can now relax the gate, e.g.:

hermes config set stt.local.no_speech_prob_threshold 0.9
hermes config set stt.local.logprob_threshold -2.0

Tests

tests/tools/test_stt_silence_hallucinations.py:

• test_confidence_thresholds_default_to_faster_whisper_values — defaults preserved.
• test_confidence_thresholds_configurable_reach_model_gate — tuned values reach the kwargs.
• test_confidence_thresholds_garbage_falls_back — bad config falls back to defaults.
• Extended test_hardened_kwargs_reach_model to assert both thresholds reach model.transcribe().

tests/tools/test_stt_silence_hallucinations.py  21 passed

Note: 3 pre-existing failures in tests/tools/test_transcription_tools.py
(test_config_device_and_compute_type_passed_to_whisper,
test_config_defaults_to_auto_when_not_set,
test_cublas_status_not_supported_retries_on_cpu) reproduce on a pristine
upstream/main checkout with none of this change — they assert the device string
passed to the WhisperModel constructor (auto/auto) and are environment-specific
(local CUDA-less resolution), unrelated to this diff, which only adds
transcribe() kwargs.

Scope

This narrows the confidence-threshold gate only; it deliberately does not touch the default-VAD path (tools/transcription_tools.py:1533), which is the other half of #74178 (VAD can strip all audio before decoding). Referencing #74178 with a non-closing keyword so that remaining VAD work stays tracked after this merges.

Refs #74178

[teknium1]

Thanks for the focused, well-covered threshold propagation fix. The current-head premise is real: build_local_transcribe_kwargs() returns without these keys at tools/transcription_tools.py:1513-1552, while _confidence_thresholds() is currently used only by the post-filter at tools/transcription_tools.py:1555-1605.

Problems

• Fixes #74178 is too broad. That issue reports a primary case where the default VAD removes all audio before decoding; current main still enables VAD at tools/transcription_tools.py:1533, and this PR intentionally does not change that path. Merging with the closing keyword would close remaining work.

Suggested changes

• Use a non-closing reference to fix: faster-whisper VAD filter + default thresholds drop non-English speech #74178, or split/narrow the issue so the unresolved VAD behavior remains tracked.

The two added kwargs otherwise fit the pinned faster-whisper API and the shared kwargs helper reaches both local transcribe calls (tools/transcription_tools.py:1642-1665).

Automated hermes-sweeper review.

[PRATHAMESH75]

Good catch on the closing keyword. This PR only threads the confidence thresholds into faster-whisper's own gate; it intentionally leaves the default-VAD path (tools/transcription_tools.py:1533) untouched, which is the other, primary half of #74178 (VAD stripping all audio before decode). Merging with Fixes would have closed that remaining work.

I've switched the PR body to a non-closing Refs #74178 and added a Scope section spelling out exactly which half this addresses, so the unresolved VAD behavior stays tracked. No code change — the threshold propagation and its tests are unchanged.

[kshitijk4poor]

Merged via #77516 — thank you @PRATHAMESH75. Your commit was cherry-picked, so you remain the author in git history.

Verified before merging: both kwargs checked against faster-whisper's own source (they gate segment retention exactly as you described), the premise re-confirmed on current main (build_local_transcribe_kwargs still omitted both), and a mutation check proved your 4 wiring tests fail without the production change. The fix makes stt.local.no_speech_prob_threshold / stt.local.logprob_threshold work end-to-end for the first time — non-English speech was the real casualty of the hardcoded internal gate. No changes to your code; the salvage exists only because your fork's base was too old for current required CI checks to run.