Quality Metrics

This README provides detailed instructions for setting up and effectively utilizing speech quality metrics.

Project Description

speech_quality_metrics is a package that offers various speech quality metrics, including MOS (Mean Opinion Score) and PESQ (Perceptual Evaluation of Speech Quality).

Installation

Requirements

Before installing the package, ensure you have the following requirements:

• Python 3.9 or higher
• pip installed

Package Installation

1. Clone the repository:

   git clone https://github.com/freds0/speech_quality_metrics
   cd speech_quality_metrics
   

2. Install the dependencies:

   pip install torch torchvision torchaudio
   pip install -e .
   

Usage

To use the speech_quality_metrics library, follow the instructions below to set up and execute various types of speech quality predictions:

Objective Metrics

For evaluating objective metrics of speech quality:

from speech_quality_metrics.objective_metrics import ObjectiveMetricsPredictor
obj_quality_predictor = ObjectiveMetricsPredictor(device)    
objective_metrics = obj_quality_predictor.predict_folder(input_dir, batch_size, search_pattern)

Parameters

• input_dir: Specify the directory containing audio files to be processed.
• batch_size: Batch size for processing multiple files in parallel.
• search_pattern: File pattern to search for in the input directory (default: *.wav).
• device: Specify the computation device (cpu or cuda).

Subjective Metrics

To assess subjective metrics of speech quality:

from speech_quality_metrics.subjective_metrics import SubjectiveMetricsPredictor
sbj_quality_predictor = SubjectiveMetricsPredictor(device)
subjective_metrics = sbj_quality_predictor.predict_folder(input_dir, batch_size, search_pattern)

Parameters

• input_dir: Specify the directory containing audio files to be processed.
• batch_size: Batch size for processing multiple files in parallel.
• search_pattern: File pattern to search for in the input directory (default: *.wav).
• device: Specify the computation device (cpu or cuda).

Intelligibility Metrics

For measuring speech intelligibility:

from speech_quality_metrics.intelligibility_metrics import IntelligibilityMetricsPredictor
int_quality_predictor = IntelligibilityMetricsPredictor(device)
intelligibility_metrics = int_quality_predictor.predict_folder(input_dir, gt_text_file, search_pattern)

Parameters

• input_dir: Specify the directory containing audio files to be processed.
• gt_text_file: Optional path to a file containing ground truth transcriptions.
• search_pattern: File pattern to search for in the input directory (default: *.wav).
• device: Specify the computation device (cpu or cuda).

Speaker Similarity Metrics

To determine speaker similarity metrics:

from speech_quality_metrics.similarity_metrics import SpeakerSimilarityPredictor
spk_quality_predictor = SpeakerSimilarityPredictor(model, device)
similarity_metrics = spk_quality_predictor.predict_folder(input_dir, gt_dir, batch_size, search_pattern)

Parameters

• input_dir: Specify the directory containing audio samples.
• gt_dir: Path to the ground truth audio file or directory.
• model: Choose from resemblyzer, ecapa_tdnn, wavlm or ecapa2.
• batch_size: Number of files to process simultaneously.
• search_pattern: File pattern to match in the directory.
• device: Computation device (cpu or cuda).

Silence Metrics

For analyzing silence within speech files:

from speech_quality_metrics.silence_metrics import SilenceMetricsPredictor
sil_quality_predictor = SilenceMetricsPredictor(threshold, interval)
silence_metrics = sil_quality_predictor.predict_folder(input_dir, search_pattern)

Parameters

• threshold: Decibel threshold for silence detection. Adjust based on your specific signal-to-noise ratio.
• interval: Interval in milliseconds for analyzing audio chunks. Smaller intervals may provide more precise measurements but require more processing time.
• input_dir: Directory containing the audio files.
• search_pattern: File pattern to match files in the directory (default: *.wav).

These examples demonstrate how to use different components of the speech_quality_metrics library to evaluate various aspects of speech quality across a collection of audio files.