Pyannote Audio

Yes. The library is MIT licensed, so you can use it in commercial products, modify it, and deploy it on your own infrastructure without restriction or royalties. The pretrained models on Hugging Face require accepting their individual terms, which permit research and commercial use. The premium pyannoteAI API is a separate paid service with higher-accuracy models.

It works without a GPU, but inference runs many times slower than real-time for typical audio files on CPU. The README benchmarks show 31 seconds of processing per hour of audio on an H100 GPU; CPU performance depends on the hardware but is substantially slower. For low-volume or offline use, CPU is fine; for production throughput, a GPU is needed.

Pyannote.audio is the open-source Python library running on your own hardware; the pyannoteAI API (Precision-2 and Precision-3 tiers) is a managed service running on pyannoteAI's servers. They share the same interface. The API offers higher accuracy and faster inference (14 seconds per hour on Precision-2 vs 31 seconds for community-1) but introduces per-call pricing and audio leaving your environment. Use the open-source library when you need data residency, cost control at volume, or the ability to fine-tune.

Pyannote.audio uses torchcodec for audio decoding, which requires FFmpeg to be installed on the system. FFmpeg supports a wide range of formats including WAV, MP3, FLAC, and MP4 audio tracks. The README documents FFmpeg as a required system dependency.

Yes. All pretrained models and pipelines are hosted on Hugging Face and can be fine-tuned on your own labeled audio data using the built-in training framework. Multi-GPU training is supported via pytorch-lightning. Fine-tuning on domain-specific recordings (call center audio, medical transcripts, broadcast content) typically reduces diarization error rates compared to the general-purpose pretrained model.