lamp: narration (pocket_tts, azelma, 24 kHz, 0.6 s gaps)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FAqjv3n4yABxvJPJHFRhVA

src/lamp/tts.py

@@ -0,0 +1,18 @@
+import json, numpy as np
+from pocket_tts import TTSModel
+S = json.load(open('script.json'))
+import pocket_tts, pathlib; m = TTSModel.load_model(config=str(pathlib.Path(pocket_tts.__file__).parent / 'config' / 'english_2026-01.yaml'))
+assert int(m.sample_rate) == 24000
+st = m.get_state_for_audio_prompt('azelma')
+gap = np.zeros(int(0.6 * 24000), dtype='<i2')
+D, chunks = [], []
+for i, t in enumerate(S):
+    parts = [(s.detach().cpu().numpy().clip(-1, 1) * 32767).astype('<i2') for s in m.generate_audio_stream(st, t, copy_state=True)]
+    a = np.concatenate(parts)
+    D.append(round(len(a) / 24000, 3)); chunks.append(a)
+    if i < len(S) - 1: chunks.append(gap)
+    print(i, D[-1], flush=True)
+full = np.concatenate(chunks)
+open('lamp.pcm', 'wb').write(full.tobytes())
+json.dump(D, open('durations.json', 'w'))
+print('total', round(len(full) / 24000, 2), flush=True)