BUG: Transcription returns only the first ~600 characters (the beginning of the audio), ignoring the rest. It doesn't matter if it's 30 seconds or 53 minutes.

[GrigK]

Izwi Bug Report Info

System

• OS: Ubuntu 24.04.4 LTS
• Kernel: Linux 6.8.0-100-generic
• CPU: 12 cores
• RAM: 31 GB (13 GB available)
• GPU: AMD Radeon RX 470/480/570/580/590 (Ellesmere)
• No NVIDIA GPU - running on CPU only

Version

• Izwi: v0.1.0-alpha-11
• CLI & Server

Steps to Reproduce

1. Start server: ./izwi-server
2. Transcribe any audio file (tested with 30 sec, 10 min, 53 min files)
3. Command: izwi transcribe <file> --language ru -f json
4. Result: Returns same short text (~600 chars) regardless of input file length

Expected Behavior

Full transcription of entire audio file.

Actual Behavior

Only returns first ~600 characters of audio, ignoring rest. Same result for:

• 30 second audio
• 10 minute audio
• 53 minute audio

Logs

Server runs fine, transcription completes, returns truncated text.

Tested Models

• Qwen3-ASR-0.6B
• Qwen3-ASR-1.7B

Both return same truncated output.

Additional Notes

• Model files download successfully (~4.7GB for 1.7B)
• Server doesn't crash on alpha-11 (unlike earlier tests)
• CPU inference works (RTF ~0.9 for 30s audio)
• The output text is always the same ~first 10-15 seconds of audio

[zinyando]

Thanks for the bug report. I'm working on adding support for longer audios and transcriptions.

I will let you know know I push support for them in a day or 2.

[zinyando]

Hey @GrigK, I tagged a new release which should fix this issue. Give it a try and let me know if it works as expected https://github.com/agentem-ai/izwi/releases/tag/v0.1.0-alpha-12

From experience, Parakeet models are faster than Qwen3-ASR...give then a try.