TTS Streaming Listening Review

Listen to the generated WAV review files, then record the human verdict that automated ffmpeg and waveform checks cannot provide.

Important: waveform sanity is PASS for these files, but that does not prove there are no audible clicks, gaps, pacing issues, or Raycast stop/cancel problems. The goal remains incomplete until a person signs off here or in the Markdown reports.

Kokoro Current

Captured from the current worktree gateway on http://127.0.0.1:45129. The sample spans five stream chunks. The committed audio files are 8-second excerpts trimmed from the full local capture to stay under the repository large-file gate.

Kokoro /tts/stream

MP3 stream decoded to WAV

waveform PASS

Duration: 8.000s
Peak / RMS: 21603 / 1643.93
Longest silence: 0.581s
Max jump: 9053

Kokoro /tts/stream/pcm

Raw PCM stream decoded to WAV

waveform PASS

Duration: 8.000s
Peak / RMS: 24207 / 1733.12
Longest silence: 0.583s
Max jump: 10091

CosyVoice3 Zero-Shot

Captured from the real local CosyVoice3 sidecar spike. Latency was not competitive with Kokoro, but these files support the quality check.

CosyVoice /tts/stream

MP3 stream decoded to WAV

waveform PASS

Duration: 2.114s
Peak / RMS: 24584 / 2226.23
Longest silence: 0.176s
Max jump: 12190

CosyVoice /tts/stream/pcm

Raw PCM stream decoded to WAV

waveform PASS

Duration: 2.440s
Peak / RMS: 18005 / 2163.26
Longest silence: 0.254s
Max jump: 10300

Human Verdict

Check these after listening to the files above and trying Option+R in Raycast against the live Kokoro gateway.

No clicks or pops at chunk boundaries No audible gaps or dropped audio between chunks Playback starts promptly for the selected-text workflow Stop Audio cancels playback during streaming Option+R replay on selected text sounds correct Kokoro should remain the local latency default

Source Reports

goals/tts-streaming-latency/listening/kokoro-current-report.md
goals/tts-streaming-latency/listening/cosyvoice3-zero-shot-m1-report.md