xiaoyuzhou-asr

Transcribe 小宇宙 podcast episodes to text using local Qwen3-ASR (Metal/CUDA accelerated).

Prerequisites

1. xyz API server running — fetches episode data and audio URLs from 小宇宙

   git clone https://github.com/ultrazg/xyz.git && cd xyz && go run .
   # Default port: 23020, change with -p
   

2. Access token — login via POST /sendCode then POST /login (see references/xyz-api.md)
3. ffmpeg — audio format conversion (brew install ffmpeg)
4. Qwen3-ASR model — download (HF Hub does NOT ship tokenizer.json):

   python3 -c "
   from huggingface_hub import snapshot_download
   snapshot_download('Qwen/Qwen3-ASR-0.6B', local_dir='models/0.6B')
   "
   

5. qwen3-asr-rs — build from source:

   git clone https://github.com/alan890104/qwen3-asr-rs.git && cd qwen3-asr-rs
   cargo build --release --example local_transcribe
   

6. tokenizer.json — auto-generated by the transcription script on first run (from vocab.json + merges.txt). No manual step needed.

Workflow

Step 1: Find Episode

TOKEN="$XYZ_ACCESS_TOKEN"
BASE="http://localhost:23020"

# Search episodes by keyword
curl -s -X POST $BASE/search \
  -H "x-jike-access-token: $TOKEN" -H "Content-Type: application/json" \
  -d '{"keyword":"关键词","type":"EPISODE"}'

# Get episode detail (contains audio URL)
curl -s -X POST $BASE/episode_detail \
  -H "x-jike-access-token: $TOKEN" -H "Content-Type: application/json" \
  -d '{"eid":"EPISODE_ID"}'

# List episodes of a podcast
curl -s -X POST $BASE/episode_list \
  -H "x-jike-access-token: $TOKEN" -H "Content-Type: application/json" \
  -d '{"pid":"PODCAST_ID","order":"desc"}'

Step 2: Download and Convert Audio

Audio URL is in data.data.media.source.url (m4a format).

mkdir -p /tmp/xiaoyuzhou-audio
curl -L -o /tmp/xiaoyuzhou-audio/episode.m4a "$AUDIO_URL"
ffmpeg -y -i /tmp/xiaoyuzhou-audio/episode.m4a -ar 16000 -ac 1 /tmp/xiaoyuzhou-audio/episode.wav

Step 3: Split Long Audio (REQUIRED for >3 min)

Podcasts are continuous speech with few silence gaps. Use fixed-interval splitting:

# Split into 3-minute segments (must split at ≥2 min for Metal GPU memory)
ffmpeg -y -i episode.wav -f segment -segment_time 180 -ar 16000 -ac 1 seg_%03d.wav

Or try silence-based splitting (may find no gaps in continuous podcasts):

ffmpeg -i episode.wav -af "silencedetect=noise=-30dB:d=2" -f null - 2>&1 | grep silence_end
ffmpeg -i episode.wav -f segment -segment_times T1,T2 -ar 16000 -ac 1 seg_%03d.wav

Step 4: Transcribe

MODEL_DIR="/path/to/models/0.6B"
ASR_BIN="qwen3-asr-rs/target/release/examples/local_transcribe"

# Transcribe each segment
for seg in seg_*.wav; do
  $ASR_BIN $MODEL_DIR $seg 2>/dev/null | grep "^Text     :" | sed 's/^Text     : //'
done

For efficiency (load model once in Rust):

use qwen3_asr::{AsrInference, TranscribeOptions, best_device};
let engine = AsrInference::load("models/0.6B", best_device())?;
for seg in segments {
    let result = engine.transcribe(&seg, TranscribeOptions::default())?;
    output.push(result.text);
}

Step 5: Format Output

Combine transcript with metadata as markdown:

# {title}
**节目**: {podcast.title} | **日期**: {pubDate} | **时长**: {duration}s

## 转录文本
{transcript}

References

• xyz API endpoints and auth: references/xyz-api.md
• Qwen3-ASR usage and performance: references/qwen3-asr.md

Token Management

• Tokens expire. If API returns 401, refresh: POST /refresh_token
• Store in env: XYZ_ACCESS_TOKEN, XYZ_REFRESH_TOKEN
• Prompt user to login if no valid token

Constraints

• MUST split audio into ≤3-minute segments for Metal GPU stability
• Audio must be WAV 16kHz mono
• tokenizer.json must be generated manually (not included in HF download)
• local_transcribe binary needed (demo binary only runs built-in test samples)
• xyz API requires Chinese phone number (+86) login
• All processing is local — audio never leaves the machine