Voice Transcription via Breeze ASR 26

This machine has a deployed local ASR service for Discord voice/audio transcription:

• Project: /home/chihmin/models-work/breeze-asr-26
• Model: MediaTek-Research/Breeze-ASR-26
• Service URL: http://127.0.0.1:8025
• systemd user unit: breeze-asr.service
• Main endpoints:
  • GET /health
  • POST /transcribe multipart upload
  • POST /transcribe_path local path form field

Quick Commands

Check service health

curl -fsS http://127.0.0.1:8025/health | python3 -m json.tool
systemctl --user --no-pager --plain status breeze-asr.service

Expected health fields include:

• ok: true
• model: MediaTek-Research/Breeze-ASR-26
• device: cuda on ROCm
• dtype: bfloat16

Transcribe an audio file by upload

curl -fsS -F "file=@/path/to/audio.ogg" http://127.0.0.1:8025/transcribe | python3 -m json.tool

Transcribe a local path directly

curl -fsS -F "path=/path/to/audio.ogg" http://127.0.0.1:8025/transcribe_path | python3 -m json.tool

Use bundled helper script

~/.pi/agent/skills/voice/scripts/voice-transcribe.sh /path/to/audio.ogg

Service Management

Start / restart / stop

systemctl --user start breeze-asr.service
systemctl --user restart breeze-asr.service
systemctl --user stop breeze-asr.service

Logs

journalctl --user -u breeze-asr.service --since '30 minutes ago' --no-pager | tail -120

Manual project scripts

/home/chihmin/models-work/breeze-asr-26/start.sh
/home/chihmin/models-work/breeze-asr-26/start-bg.sh
/home/chihmin/models-work/breeze-asr-26/stop.sh

Prefer the systemd unit for normal operation.

Discord / Pi Gateway Integration

Pi Discord gateway source:

• /home/chihmin/src/pi-discord-gateway
• Voice ASR helper: src/discord/voice-asr.ts
• Invocation integration: src/agent/invoke.ts
• Runtime service: pi-discord-gateway.service

Current behavior:

1. Discord attachment is downloaded into the channel session media folder.

2. Audio files (.ogg, .opus, .wav, .mp3, .m4a, .webm, .flac, .aac) are sent to Breeze ASR.

3. The transcription is injected into the prompt as:

   [Voice message transcription: voice-message.ogg]
   <transcribed text>
   

4. If ASR succeeds, the original audio file is not forwarded to Qwen as @file, because local Qwen cannot process raw .ogg audio and may return an empty response.

5. Non-audio attachments still go to pi as @file.

Relevant config in /home/chihmin/.config/pi-discord-gateway/config.env:

VOICE_ASR_ENABLED=true
VOICE_ASR_URL=http://127.0.0.1:8025
VOICE_ASR_TIMEOUT_MS=30000
VOICE_ASR_RETRIES=1
VOICE_ASR_RETRY_DELAY_MS=10000

After changing gateway code/config:

cd /home/chihmin/src/pi-discord-gateway
npm test
npm run lint
npm run build
systemctl --user restart pi-discord-gateway.service

Known Good Test Files

Archived Discord voice messages:

/tmp/pi-discord-files/2026-07-09/mrdlvcscyi49_voice-message.ogg
/tmp/pi-discord-files/2026-07-09/mrdlw2al91tg_voice-message.ogg
/tmp/pi-discord-files/2026-07-09/mrdlxxiwkvev_voice-message.ogg
/tmp/pi-discord-files/2026-07-09/mrdmkhlydlht_voice-message.ogg
/tmp/pi-discord-files/2026-07-09/mrdmmdl162jo_voice-message.ogg

Known transcriptions:

• mrdlvcscyi49_voice-message.ogg → 你可以試試看這個音檔嗎我現在要開始念一些東西喔
• mrdlw2al91tg_voice-message.ogg → 我是講台語 你們知道我要講什麼
• mrdmkhlydlht_voice-message.ogg → 哈囉你知道我在說什麼東西嗎
• mrdmmdl162jo_voice-message.ogg → 你知道我在做些什麼東西嗎

Performance Expectations

Observed on AMD Radeon 8060S / ROCm / bfloat16:

• 2.6s audio → about 1.7s
• 3.16s audio → about 2.6–3.8s
• 4.86s audio → about 2.5s
• 7s audio rough estimate → about 3.5–7s when idle; 7–12s+ if Qwen is busy

Gateway timeout is currently 30 seconds.

Qwen + Breeze Caveat

Qwen MTP (qwen-lcpp.service, port 8001) and Breeze ASR share the same ROCm GPU. Running both is usually possible, but heavy concurrent use can trigger ROCm instability:

ROCm error: unspecified launch failure
hipErrorLaunchFailure
CUDA error: unspecified launch failure

If this happens:

systemctl --user restart breeze-asr.service
systemctl --user restart qwen-lcpp.service

Check both services:

curl -fsS http://127.0.0.1:8025/health | python3 -m json.tool
curl -fsS http://127.0.0.1:8001/v1/models | python3 -m json.tool

Debugging Checklist

1. Is Breeze healthy?

   curl -fsS http://127.0.0.1:8025/health | python3 -m json.tool
   

2. Can the file be transcribed manually?

   ~/.pi/agent/skills/voice/scripts/voice-transcribe.sh /path/to/audio.ogg
   

3. Did Discord gateway transcribe it?

   journalctl --user -u pi-discord-gateway.service --since '10 minutes ago' --no-pager \
     | grep -Ei 'Transcribed Discord voice|Voice ASR|Attached files|empty response|Agent returned error'
   

4. If Qwen returns (empty response), verify the .ogg was not forwarded as @file after successful transcription. The fix lives in src/agent/invoke.ts.

5. If ROCm errors appear, restart affected services and consider lowering concurrent GPU load.