141 repos across 6 sub-areas
Models, datasets, and tools for automatic speech recognition (ASR), voice activity detection, speaker diarization, and related audio processing tasks. The cluster is dominated by pre-trained wav2vec2 and Conformer-based models fine-tuned on multilingual speech corpora (LibriSpeech, CommonVoice, VoxPopuli), along with supporting datasets and evaluation utilities. Most repositories are model cards and dataset artifacts rather than libraries, reflecting the practical focus on leveraging and benchmarking state-of-the-art speech models across languages and acoustic conditions.
Speech Recognition Models & ASR
39 repos
Pretrained and fine-tuned models for automatic speech recognition (ASR) and speech-to-text tasks. The cluster centers on popular open-source ASR architectures like Whisper, HuBERT, and Parakeet, along with their various pretrained checkpoints and language-specific variants. These are primarily model repositories and weights rather than training frameworks, useful for practitioners looking to integrate speech recognition capabilities into applications or fine-tune existing models for specific languages or domains.
Speech Recognition and Audio Processing
33 repos
Libraries, models, and tools for automatic speech recognition (ASR), speaker diarization, and audio processing tasks. The cluster spans multiple deployment contexts—from on-device inference using CoreML and ONNX Runtime to browser-based implementations with transformers.js—making these resources useful for building speech-enabled applications across platforms. Central repositories include optimized model variants like GPA and specialized audio models such as LFM2.5 and Raon-SpeechChat, reflecting a focus on making speech AI practical and accessible across different hardware and runtime constraints.
Speech processing and audio analysis
30 repos
Libraries, models, and tools for analyzing, segmenting, and understanding speech and speaker characteristics in audio. The cluster centers on speaker diarization (identifying who spoke when), voice activity detection, overlapped speech handling, and speaker segmentation—core tasks in speech understanding pipelines. Most repos build on or integrate with pyannote-audio, a widely-used framework for speaker-related audio analysis tasks, alongside complementary work in audio codecs and speech representation learning.
Multilingual Speech Recognition Models
21 repos
Pre-trained multilingual automatic speech recognition models built on wav2vec 2.0 and transformer architectures, enabling speech-to-text across diverse languages including Hungarian, Finnish, Persian, Chinese, Arabic, and Greek. The cluster primarily consists of fine-tuned model repositories that apply the XLSR-53 (cross-lingual speech representations) framework to language-specific datasets, leveraging PyTorch and the Hugging Face transformers library for production speech recognition applications.
Speech Recognition and Speaker Diarization
12 repos
Audio processing models and systems for automatic speech recognition (ASR), speaker identification, and speaker diarization—determining who spoke when in multi-speaker environments. The cluster includes streaming and real-time variants (sortformer, parakeet), speaker verification models (titanet), and dictionary-of-context resources (DiCoW), all built primarily in PyTorch. This is an applied machine learning area focusing on end-to-end speech understanding pipelines.
Speech Emotion Recognition with Transformers
6 repos
Deep learning models for detecting emotional attributes (valence, arousal, dominance) and categorical emotions from audio speech using transformer-based architectures like WavLM and wav2vec2. These repositories implement baseline systems and fine-tuned models for the SER Odyssey challenge and related emotion recognition tasks, leveraging PyTorch and pre-trained speech representations to classify emotional content from voice signals.