Update extract_embedding.py

[MADHUMITHASIVAKUMARR]

extract_embedding.py

Description:
This script extracts audio embeddings from a set of audio files using a pre-trained ONNX model. It processes the audio files to convert them into feature representations, which are then fed into the model to obtain embeddings. The script supports multi-threading for efficient processing of multiple audio files.

Key Features:

• Loads audio files specified in a wav.scp file and maps them to speaker identifiers from a corresponding utt2spk file.
• Resamples audio to 16 kHz if it’s not already in that format.
• Computes Mel-frequency filterbank features using the Kaldi library.
• Uses ONNX Runtime to run inference on the audio features, generating embeddings.
• Saves the resulting embeddings to specified files: utt2embedding.pt for individual utterance embeddings and spk2embedding.pt for averaged speaker embeddings.

Usage:

python extract_embedding.py --dir <directory_path> --onnx_path <onnx_model_path> [--num_thread <num_threads>]

Arguments:

• --dir: The directory containing the input files (wav.scp and utt2spk).
• --onnx_path: The path to the ONNX model file used for generating embeddings.
• --num_thread: (Optional) The number of threads to use for parallel processing. Defaults to 8.

Dependencies:

• torch: For handling tensors and saving embeddings.
• torchaudio: For loading audio files and processing them.
• onnxruntime: For running the ONNX model.
• torchaudio.compliance.kaldi: For extracting Mel-frequency features.
• tqdm: For displaying progress during processing.

Output:
The script generates two files in the specified directory:

• utt2embedding.pt: A PyTorch tensor containing embeddings for each utterance.
• spk2embedding.pt: A PyTorch tensor containing averaged embeddings for each speaker.

---