Skip to content

Latest commit

 

History

History
35 lines (30 loc) · 2.58 KB

README.md

File metadata and controls

35 lines (30 loc) · 2.58 KB

seed-tts-eval

💥 This repository contains the objective test set as proposed in our project, seed-TTS, along with the scripts for metric calculations. Due to considerations for AI safety, we will NOT be releasing the source code and model weights of seed-TTS. We invite you to experience the speech generation feature within ByteDance products. 💥

To evaluate the zero-shot speech generation ability of our model, we propose an out-of-domain objective evaluation test set. This test set consists of samples extracted from English (EN) and Mandarin (ZH) public corpora that are used to measure the model's performance on various objective metrics. Specifically, we employ 1,000 samples from the Common Voice dataset and 2,000 samples from the DiDiSpeech-2 dataset.

Requirements

To install all dependencies, run

pip3 install -r requirements.txt

Metrics

The word error rate (WER) and speaker similarity (SIM) metrics are adopted for objective evaluation.

  • For WER, we employ Whisper-large-v3 and Paraformer-zh as the automatic speech recognition (ASR) engines for English and Mandarin, respectively.
  • For SIM, we use WavLM-large fine-tuned on the speaker verification task (model link) to obtain speaker embeddings used to calculate the cosine similarity of speech samples of each test utterance against reference clips.

Dataset

You can download the test set for all tasks from this link. The test set is mainly organized using the method of meta file. The meaning of each line in the meta file: filename | the text of the prompt | the audio of the prompt | the text to be synthesized | the ground truth counterpart corresponding to the text to be synthesized (if exists). For different tasks, we adopt different meta files:

  • Zero-shot text-to-speech (TTS):
    • EN: en/meta.lst
    • ZH: zh/meta.lst
    • ZH (hard case): zh/hardcase.lst
  • Zero-shot voice conversion (VC):
    • EN: en/non_para_reconstruct_meta.lst
    • ZH: zh/non_para_reconstruct_meta.lst

Code

We also release the evaluation code for both metrics:

# WER
bash cal_wer.sh {the path of the meta file} {the directory of synthesized audio} {language: zh or en}
# SIM
bash cal_sim.sh {the path of the meta file} {the directory of synthesized audio} {path/wavlm_large_finetune.pth}