spkrec-ecapa-voxceleb

spkrec-ecapa-voxceleb

Checks whether two audio recordings come from the same speaker and can output speaker embeddings, using an ECAPA-TDNN model in SpeechBrain. For developers and speech researchers.

Made by speechbrain 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: spkrec-ecapa-voxceleb is a pretrained speaker verification model for SpeechBrain. It tells you whether two recordings come from the same person, and it can also output speaker embeddings, which are vectors of numbers that describe a voice. It was trained on the Voxceleb 1 and Voxceleb2 training data.

Who it suits: Developers and speech researchers who need to compare voices in Python are the main fit. Data teams can use the embeddings to group recordings by speaker.

Industry | Who uses it | How they use it | Example
Security | Application developers | Check whether two recordings share a speaker | Compare a new recording with one stored earlier
Audio analysis | Data teams | Extract embeddings to group recordings by voice | Sort a folder of recordings by speaker
Research | Speech researchers | Use it as a baseline for speaker verification | Compare their own results on the Voxceleb1 test set

If you need speech to text, voice cloning or a model proven on your own data, you can skip it.

Where to use it: Use it in Python scripts that run SpeechBrain.

• Python scripts with SpeechBrain and torchaudio
• Windows, macOS or Linux, with Python
• A CUDA GPU, if you pass the device option when loading

Use case | Who | What they get back
Same-speaker check | Developers | A score and a prediction of 1 for the same speaker or 0 for different ones
Voice embeddings | Data teams | One embedding per recording, ready for comparing or grouping
Baseline testing | Researchers | A pretrained reference system for speaker verification

How to set it up: You need Python, SpeechBrain and some audio.

• Python on Windows, macOS or Linux
• SpeechBrain, which the card installs from its develop branch with a pip command that points to a code repository
• torchaudio, to load your recordings
• Audio at 16kHz with one channel, or let the file-level calls convert it for you

1. Install SpeechBrain as the card describes.
2. Optionally download the model files with the Hugging Face hub package.
3. Run one of the Python snippets in the next section.

pip install -U huggingface_hub
hf download speechbrain/spkrec-ecapa-voxceleb

How to use it: Load the classifier by its model name and pass it a recording to get embeddings.

import torchaudio
from speechbrain.inference.speaker import EncoderClassifier
classifier = EncoderClassifier.from_hparams(source="speechbrain/spkrec-ecapa-voxceleb")
signal, fs =torchaudio.load('tests/samples/ASR/spk1_snt1.wav')
embeddings = classifier.encode_batch(signal)

To compare two recordings, load the verification class and call verify_files on both files.

from speechbrain.inference.speaker import SpeakerRecognition
verification = SpeakerRecognition.from_hparams(source="speechbrain/spkrec-ecapa-voxceleb", savedir="pretrained_models/spkrec-ecapa-voxceleb")
score, prediction = verification.verify_files("tests/samples/ASR/spk1_snt1.wav", "tests/samples/ASR/spk2_snt1.wav") # Different Speakers
score, prediction = verification.verify_files("tests/samples/ASR/spk1_snt1.wav", "tests/samples/ASR/spk1_snt2.wav") # Same Speaker

Tool | What it does
from_hparams | Loads the model by its name, and savedir sets where it is stored
encode_batch | Turns a loaded recording into speaker embeddings
verify_files | Compares two audio files and returns a score and a prediction
classify_file | Converts audio to the expected format when it is needed

If you use encode_batch or classify_batch, make sure your audio tensor already has the expected sampling rate.

Example: You pass verify_files two recordings from different people. It returns a score and a prediction of 0. With two recordings from the same person, the prediction is 1.

For newcomers and experts: Think of it as a voice comparer. You give it two recordings, and it says whether one person spoke in both. A good first try is to run the verification snippet on two short recordings you have the right to use.

The system is an ECAPA-TDNN model, a mix of convolutional and residual blocks. Attentive statistical pooling produces the embeddings, training used Additive Margin Softmax Loss, and verification takes the cosine distance between two embeddings. The card reports an equal error rate of 0.80% on the cleaned Voxceleb1 test set.

Load a recording at 16kHz with one channel
Convolutional and residual blocks process it
Attentive pooling makes a speaker embedding
Cosine distance compares two embeddings
A score and a 0 or 1 prediction come back

Good to know: Other people have published 6 quantized and 16 fine-tuned versions of it. The card also describes how to train the model from scratch with the SpeechBrain recipes.

• Accuracy: an equal error rate of 0.80% on the cleaned Voxceleb1 test set
• Audio format: trained on recordings at 16kHz with a single channel
• Limits: the SpeechBrain team gives no warranty on results with other datasets
• License: Apache 2.0
• Cost to run: none named
• Data it reaches: the audio files you pass in, plus the model files loaded by name from Hugging Face
• Not for: counting on its accuracy with other datasets, since no warranty is given there

On Hugging Face, 13 November 2026

• Library: speechbrain
• Downloads in the last 30 days: 1,912,322
• Likes: 370
• License: apache-2.0
• Languages: en
• Quantized versions: 6
• Fine-tunes: 16
• Tags: speechbrain, embeddings, speaker, verification, identification, pytorch, ecapa, tdnn

Links

• Model: https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb
• Files: https://huggingface.co/speechbrain/spkrec-ecapa-voxceleb/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download speechbrain/spkrec-ecapa-voxceleb

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