Summary

One-line fix: container.close() must be called before output_buffer.getvalue() in the write_chunk finalize block. The current order loses the final OGG page containing ~1-2 seconds of audio.

The Bug

When using response_format: "opus" on /v1/audio/speech, output audio is consistently truncated. The last 1-2 seconds are silently dropped. All other formats (MP3, WAV, FLAC, PCM) work correctly.

Related issue: #447

Root Cause

In api/src/services/streaming_audio_writer.py, the finalize block does:

# ❌ BEFORE (broken)
data = self.output_buffer.getvalue()  # reads buffer BEFORE final page is written
self.close()                           # closes container, writing final OGG page to buffer (too late)
return data                            # returns incomplete audio

For OGG/Opus, the container writes the final audio page to the output buffer during close(). By reading the buffer first, that last page is lost. MP3/WAV/FLAC aren't affected because their container close only writes metadata trailers, not audio frames.

Fix

# ✅ AFTER (fixed)
self.container.close()                 # writes final OGG page to buffer
data = self.output_buffer.getvalue()   # now includes all audio data
self.output_buffer.close()
return data

Test Results

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

Before fix

Note the round-number opus durations — OGG pages emit at ~1s granule boundaries, and the final partial page was being dropped.

After fix

Durations now match within ~60ms (normal codec framing overhead).

Changed Files

• api/src/services/streaming_audio_writer.py — 10 lines changed in write_chunk() finalize block

Testing

• Tested on GPU Docker build (CUDA 12.9.1, PyTorch)
• Verified with voice blending (am_puck(1)+am_liam(1)+am_onyx(0.5) at 1.2x speed)
• Confirmed MP3/WAV/FLAC output unchanged
• Sent fixed opus output as Discord voice messages — plays completely, no cutoff