speaker-diarization-community-1

Performs end-to-end speaker diarization by segmenting audio into speaker-homogeneous regions and assigning speaker labels, with explicit handling of overlapped speech regions where multiple speakers talk simultaneously. Uses a neural pipeline combining voice activity detection, speaker embedding extraction via ResNet-based encoders, and agglomerative clustering with dynamic thresholding to handle variable speaker counts and overlapping segments.

Detects speech presence/absence in audio using a neural binary classifier trained on variable-length audio frames, outputting frame-level probabilities that are post-processed with temporal smoothing and pause-duration thresholding to produce robust speech/non-speech segment boundaries. Architecture uses a ResNet-based encoder on mel-spectrogram features with attention mechanisms to handle variable audio lengths and distinguish speech from music/noise.

Extracts fixed-dimensional speaker embeddings (typically 192-512 dims) from variable-length speech segments using a ResNet-based encoder trained with metric learning objectives (e.g., AAM-Softmax, CosFace). Embeddings capture speaker identity in a learned metric space where same-speaker utterances cluster tightly and different-speaker utterances separate, enabling downstream clustering and speaker comparison without explicit speaker labels.

Orchestrates a multi-stage neural pipeline combining VAD, speaker embedding extraction, and agglomerative clustering into a single inference workflow with configurable component swapping and parameter tuning. Pipeline manages intermediate representations (mel-spectrograms, embeddings, similarity matrices) and applies post-processing (segment merging, label smoothing) to produce final speaker diarization output. Implemented as a modular PyTorch pipeline with lazy loading and batching support.

Performs hierarchical agglomerative clustering on speaker embeddings to group segments into speaker clusters, using cosine similarity as the distance metric and a dynamic threshold that adapts based on the distribution of pairwise similarities. Threshold selection uses a heuristic (e.g., elbow method, silhouette-based) to automatically determine the optimal number of speakers without requiring manual specification. Produces a dendrogram that can be cut at different levels to trade off speaker granularity.

Converts raw audio waveforms into mel-spectrogram representations (typically 80-128 mel-frequency bins, 10-25ms frame length) as input features for neural models. Includes augmentation techniques (SpecAugment, time-stretching, pitch-shifting) applied during training to improve model robustness to acoustic variability. Features are normalized per-utterance using mean-variance normalization to handle different recording conditions and microphone characteristics.

Explicitly detects and labels regions where multiple speakers overlap in time using a multi-task learning approach that jointly predicts speaker embeddings and overlap probability per frame. Overlapped regions are labeled separately from single-speaker regions, enabling downstream systems to handle them differently (e.g., separate ASR models for overlapped speech). Uses frame-level classification with temporal smoothing to produce robust overlap boundaries.

Estimates the number of distinct speakers in an audio file by analyzing the distribution of pairwise cosine similarities between speaker embeddings. Uses statistical methods (e.g., gap statistic, silhouette analysis) to identify the optimal number of clusters without requiring manual specification. Produces a confidence score for the estimated speaker count to indicate reliability.

Processes multiple audio files or long audio files in batches using streaming inference to minimize memory footprint. Divides long audio into overlapping chunks, processes each chunk independently, and merges results with overlap handling to produce seamless diarization across chunk boundaries. Supports parallel processing across multiple files with configurable batch size and GPU memory management.

Links speaker identities across multiple audio files by maintaining a speaker enrollment database of embeddings and comparing new speakers against enrolled speakers using similarity thresholding. Supports incremental enrollment (adding new speakers) and re-identification (matching speakers across files). Uses a similarity threshold to determine if a new speaker matches an enrolled speaker, with configurable sensitivity.