Opus response_format truncates audio — final 1-2 seconds lost (OGG page flush issue?)

[will-assistant]

Bug

When using response_format: "opus" on the non-streaming /v1/audio/speech endpoint, the output audio is consistently truncated by ~1-2 seconds compared to the same text rendered as MP3. The last word(s) get cut off.

Reproduction

Same text, same voice, same speed — only response_format differs:

# MP3 — full audio
curl -s http://localhost:8880/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input":"Test. The last word in this sentence should be clearly audible, and that word is Constantinople.","voice":"am_puck","model":"kokoro","response_format":"mp3","speed":1.2}' \
  --output test_mp3.mp3

# Opus — truncated
curl -s http://localhost:8880/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input":"Test. The last word in this sentence should be clearly audible, and that word is Constantinople.","voice":"am_puck","model":"kokoro","response_format":"opus","speed":1.2}' \
  --output test_opus.ogg

ffprobe -v error -show_entries format=duration -of csv=p=0 test_mp3.mp3   # → 5.016s
ffprobe -v error -show_entries format=duration -of csv=p=0 test_opus.ogg  # → 3.000s

Observations

• Opus duration is always a round integer (2.000, 3.000, 5.000, 9.000, 15.000 seconds) — suggests a page/frame boundary issue where the final OGG page isn't being flushed
• The gap scales with text length but is consistently ~1-2 seconds
• MP3, WAV, FLAC, PCM all produce full-length audio — only opus is affected
• Voice blending doesn't change the behavior — tested with single voices and blends
• MP3 → ffmpeg → OGG Opus conversion preserves the full duration as a workaround

Test matrix

Text length	MP3 duration	Opus duration	Lost
Short (1 sentence)	3.4s	2.0s	~1.4s
Medium (2 sentences)	5.0s	3.0s	~2.0s
Long (4 sentences)	10.2s	9.0s	~1.2s
Longer (6 sentences)	16.4s	15.0s	~1.4s

Environment

• Kokoro-FastAPI GPU Docker (NVIDIA, PyTorch)
• Linux host
• Tested via direct curl (not streaming)

Likely cause

The OGG Opus encoder isn't flushing the final page. The round-number durations strongly suggest the output is being truncated to a page boundary rather than including a short final page with the remaining audio.

Workaround

Generate as MP3, convert to OGG Opus with ffmpeg:

ffmpeg -i output.mp3 -c:a libopus -b:a 64k output.ogg

This preserves the full duration.