PodMind

Download, transcribe, and summarize Xiaoyuzhou FM (小宇宙) podcasts.

Quick Start

pip install -e .
cp .env.example .env          # add your DEEPSEEK_API_KEY

podmind full "https://www.xiaoyuzhoufm.com/episode/69f441cd5390b7cc928acdcc" --language zh

First run downloads the ASR model (~0.7–1.8 GB depending on backend). Subsequent runs hit the cache.

ASR Backends

Backend	Model	Size	Default	Speed (1h audio)	RTF	Output (1h)	Quality
mlx-qwen-asr	mlx-community/Qwen3-ASR-0.6B-4bit	680 MB	✅	~3 min	0.051	~12K chars	Good accuracy and punctuation, output slightly condensed
mlx-whisper	mlx-community/whisper-turbo	1.5 GB		~2.5 min	0.043	~19K chars	Fast, but may drop punctuation and hallucinate on Chinese
qwen-asr	Qwen/Qwen3-ASR-0.6B	1.8 GB		~6 min	0.105	~21K chars	Highest accuracy, most detailed output

All require Apple Silicon. qwen-asr needs torch and ffmpeg. mlx-qwen-asr needs mlx-audio.

Setup

• Python >= 3.12, Apple Silicon Mac, ffprobe (brew install ffmpeg)
• DeepSeek API key for summarization

pip install -e .              # includes the default mlx-qwen-asr backend
cp .env.example .env

Variable	Default	Description
DEEPSEEK_API_KEY	—	Required for summarization
DEEPSEEK_BASE_URL	https://api.deepseek.com	API endpoint
PODMIND_DATA_DIR	./data	Data directory

Usage

Full pipeline

podmind full "https://www.xiaoyuzhoufm.com/episode/<id>" --language zh

Step by step

podmind fetch "https://www.xiaoyuzhoufm.com/episode/<id>"
podmind transcribe <id> --language zh --profile
podmind summarize <id>

Use --asr-backend qwen-asr for highest quality. Replace <id> with the 24-char hex episode ID from the URL.

Batch

podmind batch-transcribe <id1> <id2> --language zh --profile

Benchmark

podmind bench <id> --language zh --profile
podmind bench <id> --backends qwen-asr --clip-seconds 120

Preflight check

podmind doctor
podmind doctor --asr-backend qwen-asr

CLI Reference

Flag	Commands	Description
--language	full, transcribe, batch, bench	zh, en, ja, ko, or full name (Chinese)
--asr-backend	all except fetch	mlx-qwen-asr (default), mlx-whisper, or qwen-asr
--asr-model	full, transcribe, batch	Override default model ID
--model	full, summarize	LLM model (default: deepseek-v4-pro)
--profile	full, transcribe, batch, bench	Print timing breakdown (RTF, chunk times)
--force	all	Ignore cache and re-run
--quiet	transcribe, batch	Suppress transcript output
--debug	all	Print full traceback on errors

Qwen-specific flags

Flag	Default	Description
--chunk-seconds	30	Seconds per chunk (30–1800)
--batch-size	1	Chunks per batch (1–8)
--qwen-max-new-tokens	0	Max tokens per chunk (0=auto-scale)
--qwen-dtype	float16	float16 (faster) or bfloat16

Tuning notes: reducing --chunk-seconds from 600 to 30 improved RTF from ~0.66 to ~0.11. Increasing --batch-size beyond 1 showed negligible gains on MPS — inference is the bottleneck, not per-call overhead.

Data Directory

data/
├── audio/         # Downloaded .m4a files
├── episodes/      # Scraped metadata (JSON)
├── transcripts/   # Transcription output + cache metadata
└── outputs/       # LLM summaries

Each backend writes to separate paths, so results from all backends coexist:

data/transcripts/<id>.txt                     # qwen-asr
data/transcripts/<id>.mlx-whisper.txt         # mlx-whisper
data/transcripts/<id>.mlx-qwen-asr.txt        # mlx-qwen-asr
data/outputs/<id>_summary.md                   # qwen-asr
data/outputs/<id>_summary.mlx-whisper.md       # mlx-whisper
data/outputs/<id>_summary.mlx-qwen-asr.md      # mlx-qwen-asr

Cache

Transcripts and summaries are cached with content-addressed metadata. Cache is invalidated when audio, language, model, backend, or chunk parameters change. Use --force to bypass.

Troubleshooting

ffprobe missing: brew install ffmpeg

Qwen "MPS is required": Install torch and ensure PODMIND_ASR_DEVICE=mps, or use --asr-backend mlx-whisper.

Unexpected errors: Re-run with --debug for a full traceback. Use --force to re-download corrupted audio or bypass stale cache.