DEDUPEDWho it is forA team with a backlog of calls, lectures or podcast episodes and one GPU to process them.
The problemTranscribing hours of audio one segment at a time ties up a GPU for days.
The outcomeLong files transcribed in a fraction of the time, with the same transcribe interface as before.
What you need first
- Python 3.9 or greater
- No system FFmpeg needed: audio is decoded with PyAV
- For GPU: NVIDIA cuBLAS for CUDA 12 and cuDNN 9 for CUDA 12
Steps
- Install with
pip install faster-whisperand the CUDA 12 libraries the README lists - Load
WhisperModel("turbo", device="cuda", compute_type="float16") - Wrap it with
BatchedInferencePipeline(model=model) - Transcribe with
batched_model.transcribe("audio.mp3", batch_size=16)and loop over the segments - Lower batch_size if you run out of GPU memory; the README benchmarks show batch_size=8 using more VRAM than unbatched runs
Prompt for your agent
Write a script with SYSTRAN/faster-whisper that walks <folder> of audio files and transcribes each with BatchedInferencePipeline at batch_size 16 on the turbo model, saving a .txt per file and skipping files already done. Log the time per file.
Caveat. Batching uses noticeably more VRAM, so it suits a dedicated NVIDIA GPU rather than a laptop.
Use this in your AI agent
Ask your agent
Using the Deduped MCP, get the setup steps for SYSTRAN/faster-whisper and "Clear a transcription backlog with batched GPU inference"
Add Deduped to Claude Code once
claude mcp add --transport http deduped https://dedupedai.com/mcp
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them freeShare: X LinkedIn Hacker News Email Copy link
Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer