speakrs

speakrs does speaker diarization with pyannote-level accuracy. Inference runs in Rust on the CPU, on Apple CoreML, or on NVIDIA CUDA, with no PyTorch or ONNX Runtime.

pip install speakrs

The package needs Python 3.11 or newer. One wheel covers every Python version.

Use

import speakrs

pipeline = speakrs.Pipeline.from_pretrained("cpu")  # downloads the models once
result = pipeline.diarize_file("meeting.wav")

for segment in result.segments:
    print(f"{segment.start:7.2f} {segment.end:7.2f} {segment.speaker}")

print(result.rttm())

To diarize many files, use the batch method. It decodes files ahead of inference and overlaps clustering with the next file:

results = pipeline.diarize_files(["a.wav", "b.flac", "c.mp3"])

Audio that is already in memory has to be 16 kHz mono float32. A NumPy array is read in place, without a copy:

result = pipeline.diarize_samples(samples, file_id="call-42")

The command line prints RTTM:

speakrs diarize meeting.wav --mode coreml

To install only the command line, in its own environment, use uv:

uv tool install speakrs

See API.md for the full API: the queue, async methods, configuration, progress events, cancellation, and error categories.

Example: a transcript with speakers

examples/transcribe.py combines speakrs with faster-whisper to write a transcript that says who said what. It runs diarization and transcription at the same time and gives each word to the speaker talking at that moment. uv can run it straight from GitHub:

uv run https://raw.githubusercontent.com/avencera/speakrs/master/crates/speakrs-python/examples/transcribe.py meeting.mp3

[06:31] Speaker 2: Crocodile?

[06:34] Speaker 5: Yeah, it can be a crocodile. It can be a crocodile. Firstly, it was an attempt at a T-rex, and then it sort of changed into a pelican, but it can be a crocodile.

[06:59] Speaker 1: That's lovely.

speakrs diarizes a 17-minute meeting in 0.6 seconds on an RTX 5090 and in under 5 seconds on an M4 Pro in CoreML mode, so whisper takes almost all of the time. The script picks CoreML on macOS and CUDA on Linux when it finds an NVIDIA GPU, falling back to the CPU when that GPU is older than Turing; --mode overrides that, and --whisper-model turbo gives a better transcript.

CUDA

The Linux and Windows wheels include the CUDA backend. It needs an NVIDIA driver and a Turing (RTX 20, T4) or newer GPU, but no CUDA toolkit. On a machine without a GPU, the wheel still imports and runs on the CPU. The Windows wheel passes CI but hasn't run on a GPU yet.

speakrs runs its own kernels for every layer. cuDNN 9 and cuBLAS are only a fallback for a layer that has no kernel for your GPU. If you want that fallback, install it with:

pip install "speakrs[cuda]"

The extra installs the nvidia-cudnn-cu12 and nvidia-cublas-cu12 wheels, which speakrs loads from site-packages on Linux and Windows. Leave it out if cuDNN 9 and cuBLAS 12 are already on the library path (on Windows, an absolute PATH entry or %CUDA_PATH%\bin).

Models

prepare() and from_pretrained() download the models from the avencera/speakrs-models Hugging Face repository. They go into the Hugging Face hub cache, which is set by SPEAKRS_CACHE_DIR, HF_HUB_CACHE or HF_HOME. A token in HF_TOKEN, or saved by hf auth login, is sent with downloads.

To skip the download, pass model_dir= to use a local directory instead.