transcribe-audio

name	transcribe-audio
description	ASR with ~30ms timestamp precision using Qwen3-ASR + ForcedAligner

transcribe-audio

Transcribe audio/video files with precise timestamps using Qwen3-ASR and Qwen3-ForcedAligner. Runs on CPU for maximum quality.

Models

Model	Purpose	Precision
Qwen3-ASR-1.7B	Speech recognition (52 languages)	SOTA accuracy
Qwen3-ForcedAligner-0.6B	Timestamp alignment	~30ms

Usage

# Full transcription with timestamps (default)
python skills/transcribe-audio/transcribe.py audio.wav

# Save to file
python skills/transcribe-audio/transcribe.py audio.wav --output captions.json

# Specify language (auto-detects if not specified)
python skills/transcribe-audio/transcribe.py audio.wav --language Chinese

# Fast mode (no timestamps)
python skills/transcribe-audio/transcribe.py audio.wav --no-timestamps

# Align existing text to audio (ForcedAligner only)
python skills/transcribe-audio/transcribe.py audio.wav --align-text "Your transcript text here"

Output Format

Compatible with Remotion captions.json:

{
  "segments": [
    {"text": "当", "startMs": 0, "endMs": 150},
    {"text": "全", "startMs": 150, "endMs": 280},
    {"text": "世界", "startMs": 280, "endMs": 520},
    ...
  ],
  "full_text": "当全世界都在追AI的时候...",
  "language": "Chinese",
  "model": "Qwen3-ASR-1.7B",
  "aligner": "Qwen3-ForcedAligner-0.6B"
}

Two Modes

1. Transcription Mode (default)

Transcribes audio and aligns timestamps:

python transcribe.py audio.wav

2. Alignment Mode

Use when you already have the transcript:

python transcribe.py audio.wav --align-text "已知的文字内容"

This skips ASR and uses ForcedAligner directly for ~30ms precision.

Programmatic Usage

from transcribe import transcribe_audio, transcribe_and_align

# Full transcription with timestamps
result = transcribe_audio("audio.wav", language="Chinese")
print(result["segments"])  # Word-level timestamps

# Align existing text
result = transcribe_and_align("audio.wav", text="已知的文字", language="Chinese")
print(result["segments"])  # ~30ms precision timestamps

Notes

• First run downloads model weights (~3GB for ASR, ~1GB for Aligner)
• Runs on CPU by default (quality first)
• Supports 52 languages including Chinese dialects
• ForcedAligner achieves ~30ms average absolute shift (SOTA)
• For best results, use WAV or high-quality audio