On-device voice isolation, wake words, voice activity and voiceprints. In the browser (WebAssembly) and in Python, with plugins for LiveKit Agents and Pipecat. Apache-2.0 code and weights.

Building SOTA automatic MOS

Free voices for Supertonic 3 text-to-speech: voice style JSON files with audio previews. 170+ more at superclonica.com

An agentic system for Rich Speech Annotation

Frozen EmbeddingGemma 2 speech encoder + small CTC reader for Saudi/Gulf Arabic ASR: code, public-set results, speed benchmark

Keyword Spotting System for Chinese wake word "你好真真" using Icefall & Sherpa-ONNX with multi-stage detection

Unlock FP/BF16 instruction issue rate on Nvidia 3000's cards to their professional equivalents

Ears for coding agents: let AI hear, review and make sound. CPU-only, open source.

SteerTTS: a tiny, measurably steerable English TTS. Inference code (MIT); weights at huggingface.co/WhissleAI/steer-tts

Yiddish Phonikud TTS — Hasidic Yiddish pointing, G2P and speech on BlueTTS 2.5. The Hugging Face Space.

Full-featured singing voice synthesis, running in your browser.

Formant & texture morph plugin for Botanica (petalcore) makers — VST3 / AU / Standalone, Windows & macOS. 为 Botanica 制作人打造的变形与质感插件

audiomod is a project for audio modifications, including audio manipulators such as time-stretching, pitch-shifing, formant-changing, and audio filters such as vibrato, tremolo, ring-modulation, co…

Export the STFT or ISTFT process in ONNX format.

Utilizes ONNX Runtime for TTS model.

Utilizes ONNX Runtime to transcribe audio into text.

Utilizes ONNX Runtime for audio denoising.

Unofficial reimplementation of FuseAlign: Forced Alignment in the Wild

[Experimental] Mel vocoder for singing voice: a ConvNeXt predicts gains for an analytic pulse train and glottal noise, rendered by STFT overlap-add

Sygma ASR Arabic: length-corrected N-best decoding for dialectal Arabic speech recognition (code, analysis, test audio)

Phonon TTS inference using xn

A lightweight inference server for Zipformer ASR, focused on simplifying deployment and exploring inference performance optimizations.

An 8M-parameter English text-to-speech model distilled from Kokoro-82M. 9 MB, runs 18x real time on one CPU thread.

Bud500: A Comprehensive Vietnamese ASR Dataset

Repository for the paper : Synthetic Speech Augmentation for Low-Resource Estonian and Slovenian ASR: Comparing Parakeet-TDT and Whisper.

LiveKit End-of-Turn Benchmark

Ming-3.0-tts is an open-source TTS model family with 0.5B dense and 7.9B-A1.3B MoE variants, supporting multilingual speech, voice cloning, voice design, and natural-language control.