metadata

license: apache-2.0
tags:
  - audio
  - speech
  - speaker
  - speaker-recognition
  - speaker-embedding
  - speaker-verification
  - speaker-identification
  - speaker-encoder
  - tflite
  - voice
library_name: sidlingvo

Conformer based multilingual speaker encoder

Summary

This is a massively multilingual conformer-based speaker recognition model.

The model was trained with public data only, using the GE2E loss.

Papers:

Multilingual: https://arxiv.org/abs/2104.02125
GE2E loss: https://arxiv.org/abs/1710.10467

@inproceedings{chojnacka2021speakerstew,
  title={{SpeakerStew: Scaling to many languages with a triaged multilingual text-dependent and text-independent speaker verification system}},
  author={Chojnacka, Roza and Pelecanos, Jason and Wang, Quan and Moreno, Ignacio Lopez},
  booktitle={Prod. Interspeech},
  pages={1064--1068},
  year={2021},
  doi={10.21437/Interspeech.2021-646},
  issn={2958-1796},
}

@inproceedings{wan2018generalized,
  title={Generalized end-to-end loss for speaker verification},
  author={Wan, Li and Wang, Quan and Papir, Alan and Moreno, Ignacio Lopez},
  booktitle={International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  pages={4879--4883},
  year={2018},
  organization={IEEE}
}

Usage

Run use this model, you will need to use the siglingvo library: https://github.com/google/speaker-id/tree/master/lingvo

Since lingvo does not support Python 3.11 yet, make sure your Python is up to 3.10.

Install the library:

pip install sidlingvo

Example usage:

import os
from sidlingvo import wav_to_dvector
from huggingface_hub import hf_hub_download

repo_id = "tflite-hub/conformer-speaker-encoder"
model_path = "models"
hf_hub_download(repo_id=repo_id, filename="vad_long_model.tflite", local_dir=model_path)
hf_hub_download(repo_id=repo_id, filename="vad_long_mean_stddev.csv", local_dir=model_path)
hf_hub_download(repo_id=repo_id, filename="conformer_tisid_medium.tflite", local_dir=model_path)

enroll_wav_files = ["your_first_wav_file.wav"]
test_wav_file = "your_second_wav_file.wav"
runner = wav_to_dvector.WavToDvectorRunner(
    vad_model_file=os.path.join(model_path, "vad_long_model.tflite"),
    vad_mean_stddev_file=os.path.join(model_path, "vad_long_mean_stddev.csv"),
    tisid_model_file=os.path.join(model_path, "conformer_tisid_medium.tflite"))
score = runner.compute_score(enroll_wav_files, test_wav_file)
print("Speaker similarity score:", score)