audiopipe
Pipeline for audio background noise removal, speaker diarization, and transcription.
Process multi-speaker recordings with Demucs, pyannote, and Whisper locally.
AudioPipe runs audio separation, speaker diarization, and transcription in a single local pipeline. It outputs structured JSON files containing transcripts, speaker labels, and timestamps for multi-speaker recordings.
What it is
The tool combines Demucs for vocal extraction, pyannote for speaker diarization, and Whisper for speech-to-text. It handles background noise removal, maps speakers to chronological segments, and supports chunked processing for long audio files.
Who it's for
It suits professionals needing structured transcripts with speaker labels from clear multi-speaker audio. It does not suit those working with poor quality audio or requiring a graphical user interface.
Key features
Audio Source Separation
Extracts vocals from background music or noise using Demucs.
Speaker Diarization
Identifies and separates different speakers with optional Hugging Face token integration.
Chunked Transcription
Transcribes long audio files in fixed-length pieces with globally consistent speaker labels.
Step Skipping
Allows starting or resuming the pipeline from separation, diarization, or transcription steps.
Up and running in 3 steps
Our honest verdict
What's great
- Outputs structured JSON with timestamps, text, and speaker attribution
- Supports step skipping and resuming via intermediate file detection
- Provides Apple Silicon MPS support alongside CUDA and CPU modes
- Handles long recordings over an hour using chunked transcription
Where it falls short
- Requires installing Python, FFmpeg, and Python packages manually
- Needs a CUDA-compatible GPU or Apple Silicon for reasonable processing speeds
- Optional enhanced diarization accuracy requires a Hugging Face account and token
Pricing & license
The only cost is running it yourself on your own hardware, utilizing local compute resources.
Open remote AI-training roles
880 open remote AI-training roles, updated November 15, 2026; most pay $20–$175 an hour. Zekai lists them free; the application goes to the employer’s platform.
Alternatives
Use it, or have us set it up
FAQ
Can I run this without a GPU?
Yes, CPU mode is available, though a CUDA-compatible GPU or Apple Silicon is recommended for performance.
How do I process files longer than an hour?
Use the --chop and --chunk-minutes flags to transcribe long audio files in fixed-length pieces.
