personal memory agent
observe/transcribe: restore CMN and snip_edges=True in wespeaker fbank front-end
Phase 2 calibration measured 16.39% EER on VoxCeleb1-O (22.7x worse than
the published 0.723%) with snip_edges=False and no CMN. Restore the
WeSpeaker training-time convention: snip_edges=True framing plus
per-utterance cepstral mean normalization, matching the POC reference
at scratch/wespeaker-poc/wespeaker_encoder.py.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
+4
-2
+4
-2
observe/transcribe/main.py
+4
-2
observe/transcribe/main.py
···
170
170
opts = knf.FbankOptions()
171
171
opts.frame_opts.samp_freq = float(sample_rate)
172
172
opts.frame_opts.dither = 0.0
173
-
opts.frame_opts.snip_edges = False
173
+
opts.frame_opts.snip_edges = True
174
174
opts.frame_opts.frame_length_ms = 25.0
175
175
opts.frame_opts.frame_shift_ms = 10.0
176
176
opts.mel_opts.num_bins = 80
···
186
186
if not frames:
187
187
return np.zeros((0, 80), dtype=np.float32)
188
188
189
-
return np.stack(frames, axis=0).astype(np.float32)
189
+
feats = np.stack(frames, axis=0).astype(np.float32)
190
+
feats = feats - feats.mean(axis=0, keepdims=True)
191
+
return feats
190
192
191
193
192
194
def _get_jsonl_path(audio_path: Path) -> Path: