[Bugfix] Validate Breeze PCM speech content and calibrate HNR guard (#8044)

Signed-off-by: Yueqian Lin <linyueqian@outlook.com>
This commit is contained in:
Yueqian Lin authored and GitHub committed 2026-09-24 07:29:22 +00:00
1 parent dbd6a35dd2
commit 1d77e96d4d
4 files changed
+216 -12

No files matched your search

+39 -9
View File
@@ -688,7 +688,9 @@ def _compute_pcm_hnr_db(pcm_samples: np.ndarray, sr: int = _PCM_SPEECH_SAMPLE_RA
peak = float(np.max(ac[min_lag:max_lag]))
if 0 < peak < 1:
hnr_values.append(10 * np.log10(peak / (1 - peak + 1e-10)))
return float(np.mean(hnr_values)) if hnr_values else 0.0
# Silence and constant signals provide no usable harmonic estimate. They
# must not pass a model-specific floor of zero or below.
return float(np.mean(hnr_values)) if hnr_values else float("-inf")
def _assert_pcm_int16_speech_hnr(
@@ -714,10 +716,7 @@ def _assert_pcm_int16_speech_hnr(
pcm_samples = np.frombuffer(audio_bytes, dtype=np.int16).astype(np.float32) / 32768.0
hnr = _compute_pcm_hnr_db(pcm_samples, sr=sr)
print(f"PCM speech HNR: {hnr:.2f} dB (threshold: {min_hnr_db} dB, sr={sr})")
assert hnr >= min_hnr_db, (
f"Audio distortion detected: HNR={hnr:.2f} dB < {min_hnr_db} dB. "
"Voice clone decoder may be losing ref_code speaker context on later chunks."
)
assert hnr >= min_hnr_db, f"PCM speech HNR={hnr:.2f} dB is below the configured floor of {min_hnr_db} dB."
def _response_has_audio_output(response: Any) -> bool:
@@ -754,11 +753,38 @@ def _omni_assertion_needs_audio_transcript(request_config: dict[str, Any], run_l
def _speech_assertion_needs_audio_transcript(request_config: dict[str, Any], run_level: str | None) -> bool:
if run_level not in {"advanced_model", "full_model"}:
return False
if request_config.get("response_format") == "pcm":
if request_config.get("response_format") == "pcm" and "transcript_pcm_sample_rate" not in request_config:
return False
return bool(request_config.get("transcript_expected_text", request_config.get("input")))
def _speech_pcm_sample_rate(request_config: dict[str, Any]) -> int:
expected = request_config.get("expected_sample_rate")
if "transcript_pcm_sample_rate" in request_config:
sample_rate = request_config["transcript_pcm_sample_rate"]
assert type(sample_rate) is int and sample_rate > 0, "transcript_pcm_sample_rate must be a positive integer"
assert expected is None or int(expected) == sample_rate, (
"transcript_pcm_sample_rate must match expected_sample_rate"
)
return sample_rate
return int(expected or _PCM_SPEECH_SAMPLE_RATE_HZ)
def _speech_audio_for_transcription(audio_bytes: bytes | None, request_config: dict[str, Any]) -> bytes | None:
"""Give ASR a WAV container when a speech test opts into raw PCM transcription."""
if not audio_bytes or request_config.get("response_format") != "pcm":
return audio_bytes
sample_rate = _speech_pcm_sample_rate(request_config)
assert len(audio_bytes) % 2 == 0, "PCM byte length must be aligned to int16"
with io.BytesIO() as buffer:
with wave.open(buffer, "wb") as wav_file:
wav_file.setnchannels(1)
wav_file.setsampwidth(2)
wav_file.setframerate(sample_rate)
wav_file.writeframes(audio_bytes)
return buffer.getvalue()
def _resolve_audio_transcript(
response: Any,
request_config: dict[str, Any],
@@ -777,7 +803,11 @@ def _resolve_audio_transcript(
existing = getattr(response, "audio_content", None)
if isinstance(existing, str) and existing.strip():
return existing
audio_bytes = getattr(response, "audio_bytes", None)
audio_bytes = (
_speech_audio_for_transcription(getattr(response, "audio_bytes", None), request_config)
if speech_api
else getattr(response, "audio_bytes", None)
)
if not audio_bytes:
return None
return convert_audio_bytes_to_text(
@@ -1005,7 +1035,7 @@ def assert_audio_speech_response(response: Any, request_config: dict[str, Any],
_assert_pcm_int16_speech_hnr(
response.audio_bytes,
min_hnr_db=min_hnr_db,
sr=int(request_config.get("expected_sample_rate") or _PCM_SPEECH_SAMPLE_RATE_HZ),
sr=_speech_pcm_sample_rate(request_config),
)
transcript = _resolve_audio_transcript(response, request_config, run_level, speech_api=True)
@@ -1016,7 +1046,7 @@ def assert_audio_speech_response(response: Any, request_config: dict[str, Any],
print(f"input text is: {expected_text}")
_assert_transcript_matches(
transcript,
getattr(response, "audio_bytes", None),
_speech_audio_for_transcription(getattr(response, "audio_bytes", None), request_config),
expected_text,
threshold=0.9,
escalation_model=request_config.get("transcript_escalation_model"),