paraclap

This repo includes the official PyTorch checkpoint of ParaCLAP – Towards a general language-audio model for computational paralinguistic tasks

Abstract

Contrastive language-audio pretraining (CLAP) has recently emerged as a method for making audio analysis more generalisable. Specifically, CLAP-style models are able to ‘answer’ a diverse set of language queries, extending the capabilities of audio models beyond a closed set of labels. However, CLAP relies on a large set of (audio, query) pairs for pretraining. While such sets are available for general audio tasks, like captioning or sound event detection, there are no datasets with matched audio and text queries for computational paralinguistic (CP) tasks. As a result, the community relies on generic CLAP models trained for general audio with limited success. In the present study, we explore training considerations for ParaCLAP, a CLAP-style model suited to CP, including a novel process for creating audio-language queries. We demonstrate its effectiveness on a set of computational paralinguistic tasks, where it is shown to surpass the performance of open-source state-of-the-art models.

Instruction

Before Evaluation, I would recommand to clone the repo from HuggingFace or GitHub

Evaluation

import os
import torch
import librosa
from transformers import logging
from transformers import AutoTokenizer
from models_xin import CLAP
from utils import compute_similarity


if __name__ == '__main__':
    logging.set_verbosity_error()
    ckpt = torch.hub.load_state_dict_from_url(
            url="https://huggingface.co/KeiKinn/paraclap/resolve/main/best.pth.tar?download=true",
            map_location="cpu",
            check_hash=True,
        )
    
    text_model = 'bert-base-uncased'
    audio_model = 'audeering/wav2vec2-large-robust-12-ft-emotion-msp-dim'
    
    device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
    
    candidates = ['happy', 'sad', 'surprise', 'angry'] # free to adapt it to your need
    wavpath = '[Waveform path]' # single channel wavform

    waveform, sample_rate = librosa.load(wavpath, sr=16000)
    x = torch.Tensor(waveform)

    tokenizer = AutoTokenizer.from_pretrained(text_model)

    candidate_tokens = tokenizer.batch_encode_plus(
        candidates,
        padding=True,
        truncation=True,
        return_tensors='pt'
    )

    model = CLAP(
        speech_name=audio_model,
        text_name=text_model,
        embedding_dim=768,
    )

    model.load_state_dict(ckpt)
    model.to(device)
    print(f'Checkpoint is loaded')
    model.eval()

    with torch.no_grad():
        z = model(
            x.unsqueeze(0).to(device),
            candidate_tokens
        )

    similarity = compute_similarity(z[2], z[0], z[1])
    prediction = similarity.T.argmax(dim=1)
    
    result = candidates[prediction]

Citation Info

ParaCLAP has been accept at InterSpeech 2024 for presentation.

@inproceedings{Jing24_PTA,
  title     = {ParaCLAP – Towards a general language-audio model for computational paralinguistic tasks},
  author    = {Xin Jing and Andreas Triantafyllopoulos and Björn Schuller},
  year      = {2024},
  booktitle = {Interspeech 2024},
  pages     = {1155--1159},
  doi       = {10.21437/Interspeech.2024-1315},
  issn      = {2958-1796},
}

---

license: cc-by-nc-nd-4.0 language:

• en

---