qwen-transkrib

Multilingual audio transcription with speaker diarization. Supports Qwen3-ASR and GigaAM v3 (e2e_rnnt) (Russian ASR, WER 2.76%) backends. Uses pyannote for speaker identification and language-specific punctuation restoration.

Quickstart

pip install uv
git clone https://github.com/GRomR1/qwen-transkrib.git
cd qwen-transkrib
uv sync
./apply_patches.sh  # Required for MetaX GPU
uv run qwen-transkrib transcribe audio.wav --context "Google, Microsoft, Amazon"

Requirements

• Python 3.12
• GPU: MetaX C500/C550/C650 or NVIDIA CUDA
• RAM: 16 GB minimum, 32 GB recommended
• Disk: 10 GB for models

Installation

uv (recommended)

pip install uv
uv sync
./apply_patches.sh  # Apply MetaX-specific patches

MetaX GPU

On MetaX (MACA) GPUs, PyTorch links against libmctlassEx.so from the MACA SDK, which uv sync does not install. Install it via apt before running — the version must match your installed MACA SDK (3.8.1.3 for the C550):

sudo apt-get install -y mctlassex_3.8.1=3.8.1.3

Without it, import torch fails with ImportError: libmctlassEx.so: cannot open shared object file. Note that the similarly named mctlass_3.8.1 package is headers-only and does not provide this library.

pip

pip install .
# Apply patches manually if on MetaX

Docker

docker build -t qwen-transkrib .
docker run --gpus all -v $(pwd)/data:/data qwen-transkrib transcribe /data/audio.wav

# Or use Docker Compose
docker compose up asr

Usage

CLI

# Basic transcription (Russian)
uv run qwen-transkrib transcribe audio.wav

# English transcription
uv run qwen-transkrib transcribe audio.wav --language English

# With diarization and context for term correction
uv run qwen-transkrib transcribe audio.wav --context "Google, Microsoft, Amazon"

# All options
uv run qwen-transkrib transcribe input.webm \
  --output-dir ./out \
  --language Russian \
  --format srt,json,txt \
  --context "Google, Microsoft, Amazon" \
  --device cuda:0 \
  --backend gigaam \
  --punct \
  --vad \
  --normalize \
  --glossary "гугл=Google,майкрософт=Microsoft"

# Benchmark WER against reference dataset
uv run qwen-transkrib bench bond005/podlodka_speech -n 20 --backend gigaam
uv run qwen-transkrib bench bond005/sberdevices_golos_10h_crowd -n 50 --backend gigaam

# Show environment info
uv run qwen-transkrib info

Python API

from pathlib import Path
from qwen_transkrib import transcribe_file, diarize_file, assign_speakers, Settings

settings = Settings()
words, text, lang = transcribe_file(Path("audio.wav"), settings, context="Google, Microsoft")
diar = diarize_file(Path("audio.wav"), settings)
segments = assign_speakers(words, diar)

Supported Languages

Configuration

Environment Variables

# Required for diarization (HuggingFace token)
export HF_TOKEN=your_token_here

# Optional: Model settings
export QWEN_ASR_MODEL=Qwen/Qwen3-ASR-1.7B
export QWEN_ALIGNER_MODEL=Qwen/Qwen3-ForcedAligner-0.6B
export QWEN_LANGUAGE=Russian
export QWEN_DEVICE=cuda:0

.env file

Create .env in project root:

HF_TOKEN=your_token_here

Output Formats

• SRT - SubRip subtitles (for video players)
• VTT - WebVTT subtitles (for web)
• JSON - Full transcription with word-level timestamps
• TXT - Plain text

Documentation

• Architecture — data flow, modules, design decisions
• Benchmarks — WER, RT factor, memory usage
• Troubleshooting — common issues and fixes

Development

# Install dev dependencies
uv sync --group dev

# Run tests
uv run pytest tests/ -v

# Lint
uv run ruff check src/

# Type check
uv run mypy src/

License

Apache-2.0