wespeaker-voxceleb-resnet34-LM

wespeaker-voxceleb-resnet34-LM

Turns a voice recording into a speaker embedding with a pyannote.audio wrapper around the WeSpeaker ResNet34 model, so you can compare how alike two speakers sound. For speech and audio developers.

Made by pyannote on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is a pyannote.audio wrapper around the WeSpeaker ResNet34 speaker embedding model trained on VoxCeleb. It converts a stretch of speech into a vector that describes who is speaking. You can compare two vectors to see how alike two voices are.

The card says it needs pyannote.audio version 3.1 or higher.

Who it suits: It suits developers and researchers who build speaker verification, diarization or voice search tools and want a ready speaker embedding model inside pyannote.audio.

Industry | Who uses it | How they use it | Example
Speech technology | Developers | Compare two recordings for the same speaker | Check if two clips share a voice
Meetings | Product teams | Label who spoke in a recording | Group segments by speaker
Research | Speech researchers | Extract embeddings for experiments | Cluster speakers in a dataset
Media | Audio engineers | Find segments from one speaker in long audio | Pull one guest's lines

If you want a full diarization pipeline in one step, or speech recognition, you can skip it, since this model only produces embeddings. Use it only on audio you have the right to process.

Where to use it: The model loads through pyannote.audio in Python, which runs on Windows, macOS and Linux, on a CPU or a CUDA GPU.

• Python scripts that use pyannote.audio
• A CUDA GPU, for faster extraction
• Whole-file, excerpt or sliding-window embedding extraction

Use case | Who | What they get back
Speaker comparison | Developers | A cosine distance between two voices
Excerpt embedding | Researchers | A vector for a chosen time range
Sliding window | Diarization builders | One vector per window along the audio

How to set it up: You need Python with pip and pyannote.audio at version 3.1 or newer. The card shows loading the model by its Hugging Face id.

• Python with pip
• pyannote.audio 3.1 or newer
• scipy, if you want the cosine distance example

1. Install the Hugging Face hub package.
2. Download the model files, or load them by id in code.
3. Create an Inference object for the model.

pip install -U huggingface_hub
hf download pyannote/wespeaker-voxceleb-resnet34-LM

How to use it: You load the model, wrap it in an Inference object and pass it audio files. With the whole-file window you get one vector per file, and you compare two with a cosine distance.

# instantiate pretrained model
from pyannote.audio import Model
model = Model.from_pretrained("pyannote/wespeaker-voxceleb-resnet34-LM")

from pyannote.audio import Inference
inference = Inference(model, window="whole")
embedding1 = inference("speaker1.wav")
embedding2 = inference("speaker2.wav")
# `embeddingX` is (1 x D) numpy array extracted from the file as a whole.

from scipy.spatial.distance import cdist
distance = cdist(embedding1, embedding2, metric="cosine")[0,0]
# `distance` is a `float` describing how dissimilar speakers 1 and 2 are.

Tool | What it does
Model.from_pretrained | Loads the speaker embedding model
Inference with a whole window | Gives one embedding for a whole file
Inference with a sliding window | Gives one embedding per window
inference.crop | Gives an embedding for an excerpt

Example: You pass two WAV files to the Inference object and get one vector for each. The cosine distance between them is a number, and a lower number means the voices sound more alike.

For newcomers and experts: If you are new, run the two-file example with a pair of short clips from the same person and then a pair from different people. Compare the two distances.

If you know the field, the sliding-window mode returns an N by D feature object, where each row covers a window of a set duration and step. That fits diarization pipelines that cluster these windows by speaker.

Load the model by its id
Wrap it in an Inference object
Pass audio or an excerpt
Compare or cluster the vectors

Good to know: The model follows the license of its training data, VoxCeleb, so it is under Creative Commons Attribution 4.0. The card also notes a hosted option for production use.

• License: cc-by-4.0
• Cost to run: none named for this model
• Data it reaches: the audio files you pass in, and the model files you download
• Not for: use on voices you have no right to process

On Hugging Face, 13 November 2026

• Library: pyannote-audio
• Downloads in the last 30 days: 5,889,758
• Likes: 226
• License: cc-by-4.0
• Quantized versions: 2
• Fine-tunes: 9
• Tags: pyannote-audio, pytorch, pyannote, pyannote-audio-model, wespeaker, audio, voice, speech, speaker, speaker-recognition, speaker-verification, speaker-identification

Links

• Model: https://huggingface.co/pyannote/wespeaker-voxceleb-resnet34-LM
• Files: https://huggingface.co/pyannote/wespeaker-voxceleb-resnet34-LM/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download pyannote/wespeaker-voxceleb-resnet34-LM

Then follow the usage in its model card. Nothing in this kit runs by itself.