16 models · ranked by HuggingFace downloads
Kokoro-82M is a compact 82-million-parameter text-to-speech model fine-tuned from StyleTTS2, targeting natural-sounding English speech synthesis at a size runnable on CPU or modest GPU. Released under Apache 2.0 with a HuggingFace DOI, it gained attention as a high-quality open TTS model at significantly smaller scale than most alternatives. It supports multiple English voice styles.
XTTS-v2 is Coqui's multilingual text-to-speech model supporting 17 languages with voice cloning from a short audio sample. It uses a GPT-style decoder for speech token generation, enabling zero-shot speaker cloning without fine-tuning. The model was released before Coqui's closure and remains available under a non-standard license.
Qwen3-TTS-12Hz-1.7B-CustomVoice synthesizes speech waveforms from text input. It produces natural-sounding audio and supports different speaking rates or voice styles depending on the variant.
Chatterbox is Resemble AI's open-source text-to-speech model offering voice cloning and expressive speech synthesis. It is designed as a production-grade TTS system with controllable prosody and emotion.
OmniVoice from k2-fsa is a multilingual speech model targeting end-to-end ASR and voice processing tasks. Published as part of the k2/Lhotse/sherpa-onnx ecosystem for server and edge speech applications.
Qwen3-TTS CustomVoice is the 0.6B variant of Qwen's TTS family focused on voice customization from reference audio. At 12Hz token rate and 0.6B parameters, it's designed for constrained environments where a full 1.7B TTS model is too heavy. Supports 9 languages including CJK languages and major European languages.
indic-parler-tts is a TTS model that generates audio directly from text tokens, enabling low-latency speech synthesis without a separate vocoder stage.
Qwen3-TTS-12Hz-0.6B-Base is a TTS model that generates audio directly from text tokens, enabling low-latency speech synthesis without a separate vocoder stage.
VibeVoice-Realtime-0.5B synthesizes speech waveforms from text input. It produces natural-sounding audio and supports different speaking rates or voice styles depending on the variant.
Qwen3-TTS VoiceDesign is a 1.7B text-to-speech model operating at 12Hz token rate, designed to support custom voice creation alongside standard TTS. It covers multiple languages and generates expressive speech from text input. Apache-2.0 licensed and part of Qwen's audio model family.
F5-TTS synthesizes speech waveforms from text input. It produces natural-sounding audio and supports different speaking rates or voice styles depending on the variant.
MOSS-TTS is OpenMOSS's multilingual text-to-speech model supporting 20 languages including Chinese, English, German, Japanese, Korean, Russian, and Hebrew. It uses a delay-based autoregressive architecture (moss_tts_delay) for high-quality speech synthesis with natural prosody. Apache-2.0 licensing makes it a viable open alternative to commercial TTS APIs for multilingual applications.
Kokoro-82M is a lightweight 82M-parameter text-to-speech model converted to ONNX by the HuggingFace ONNX community, enabling browser-based and edge TTS via Transformers.js. It uses the StyleTTS2 architecture, which separates style and content representations to produce expressive speech without large acoustic models. The ONNX conversion allows direct client-side inference without a server.
MMS-TTS-HAT is Meta's Massively Multilingual Speech TTS model for Haitian Creole (hat), part of the MMS project targeting 1000+ languages. It uses VITS architecture for end-to-end speech synthesis. CC-BY-NC-4.0 licensed — non-commercial use only.
VoxCPM2 is an open-source text-to-speech model available on HuggingFace. Details are sourced from the public model registry.
Kokoro-82M-bf16 synthesizes speech waveforms from text input. It produces natural-sounding audio and supports different speaking rates or voice styles depending on the variant.