[

Open Source

]

Open-Source Voice AI Models

109 open-source STT, TTS, speech-to-speech, VAD and agent-framework projects for voice AI founders and engineers, with licences and real builder feedback.

Updated October 2026 · Curated from builder discussions in the Unio community and 2026 benchmarks

Submit your open-source model

Browse all models

LAST UPDATED: OCTOBER 2026

FILTER BY LICENCE: SEE WHAT YOU CAN SHIP COMMERCIALLY

Explore every open-source voice AI model

Filter by layer of the stack, licence and Indian-language support. Each card shows what the model does, where it shines, what to watch out for and what builders say about it.

Showing 0 of 0 models · GitHub stars and Hugging Face downloads (last 30 days) checked October 2026

No models match these filters.

The shortlist

What voice AI builders actually use

Ranked by how often builders in the Unio community recommend each project, and how positively they talk about it.

0open models and tools tracked

0have licences that allow commercial use

0layers of the voice stack covered

0built for or tested on Indian languages

Community top 10

Most used by builders worldwide

Best pick for each job

Speech-to-textfaster-whisperThe default STT; 4× faster than Whisper

TTS on CPUKokoro82M params, under 80ms, cheapest to run

Expressive TTSOrpheusFine-tune your own voice (also Qwen3-TTS)

Voice activitySilero VADTiny, accurate, runs on CPU

Agent frameworkLiveKit AgentsFree open SFU; or Pipecat for integrations

Noise suppressionRNNoiseLight default for SIP; try DeepFilterNet

Speech-to-speechMoshiOpen full-duplex reference; PersonaPlex for 200ms

Indian languagesAI4Bharat IndicConformerAll 22 scheduled languages, MIT

Unio · India’s Voice AI communityuniocommunity.com/open-source-voice-ai-models · Updated October 2026

Still unsolved

Gaps builders keep hitting

Problems the community raises again and again that no open-source project has cracked yet. Good places to build.

Indian-language TTS worth shipping

Builders are still hunting for Hinglish-capable open TTS. The Orpheus vs Qwen3-TTS debate is unresolved, and XTTS-v2’s licence blocks commercial use.

A VAD that beats Silero

Asked often, never answered. Power users pair Silero with Pipecat SmartTurn for better turn detection instead.

Filtering background human speech

Removing other people talking on a call is still Krisp and ai-coustics territory. There is no open-source winner yet.

Commercially usable open diarization

pyannote has the best quality, but its pretrained weights need a paid licence for commercial use. NeMo Sortformer caps at 4 speakers.

Open full-duplex for Indian languages

Moshi, PersonaPlex, Nemotron VoiceChat and MiniCPM-o are all English- or Chinese-first. No open full-duplex Indic model exists yet.

Licence traps

Piper moved from MIT to GPL-3.0, and XTTS-v2, F5-TTS and several top Arena models are non-commercial. Always check before you ship.

Unio · India’s Voice AI communityuniocommunity.com/open-source-voice-ai-models · Updated October 2026

OpenAI Whisper (large-v3 / Turbo): Multilingual transcription, 99 languages (Speech-to-Text, licence MIT)

faster-whisper: CTranslate2 re-implementation of Whisper (Speech-to-Text, licence MIT)

whisper.cpp: Whisper in C/C++ (Speech-to-Text, licence MIT)

WhisperX: Whisper + word alignment + diarization (Speech-to-Text, licence BSD-2-Clause)

NVIDIA Parakeet (TDT 0.6B v3): Fast transducer ASR (Speech-to-Text, licence CC-BY-4.0 (weights); NeMo toolkit Apache-2.0)

NVIDIA Canary (1B v2 / Qwen 2.5B): Multilingual ASR + speech translation (Speech-to-Text, licence CC-BY-4.0)

NVIDIA NeMo toolkit: Full ASR/TTS/NLP toolkit (Speech-to-Text, licence Apache-2.0)

Voxtral (Mistral): Speech understanding model (Speech-to-Text, licence Apache-2.0)

Qwen3-ASR (1.7B / 0.6B): Multilingual ASR (Speech-to-Text, licence Apache-2.0)

SenseVoice (Alibaba FunASR): Multilingual ASR (Speech-to-Text, licence Apache-2.0)

Moonshine: Tiny on-device STT (Speech-to-Text, licence MIT (some legacy models non-commercial))

Vosk: Offline lightweight STT (Speech-to-Text, licence Apache-2.0)

Sherpa-ONNX: Streaming ASR toolkit (Zipformer etc.) (Speech-to-Text, licence Apache-2.0)

Wav2Vec2 / SpeechBrain / ESPnet: Research ASR toolkits (Speech-to-Text, licence Apache-2.0)

IBM Granite Speech: Enterprise ASR (Speech-to-Text, licence Apache-2.0)

Kokoro-82M: Lightweight multilingual TTS (Text-to-Speech, licence Apache-2.0)

Piper: Fast neural TTS (VITS/ONNX) (Text-to-Speech, licence GPL-3.0 (current; original MIT repo archived Oct 2025))

Orpheus (Canopy Labs): Expressive LLM-based TTS (Text-to-Speech, licence Apache-2.0)

Qwen3-TTS (0.6B/1.7B): LLM-based TTS + cloning variant (Text-to-Speech, licence Apache-2.0)

Chatterbox (Resemble AI): Zero-shot voice cloning TTS (Text-to-Speech, licence MIT (watermark always on))

F5-TTS: Flow-matching zero-shot TTS (Text-to-Speech, licence CC-BY-NC (weights — non-commercial))

XTTS-v2 (Coqui): Multilingual voice cloning (Text-to-Speech, licence CPML — non-commercial)

Zonos (Zyphra): 1.6B expressive TTS (Text-to-Speech, licence Apache-2.0)

Dia (Nari Labs): 1.6B dialogue TTS (Text-to-Speech, licence Apache-2.0)

Sesame CSM-1B: Conversational speech model (Text-to-Speech, licence Apache-2.0)

Fish Speech / OpenAudio S1: Multilingual cloning TTS (Text-to-Speech, licence Apache-2.0)

CosyVoice 3.0 (Alibaba): Multilingual cloning TTS (Text-to-Speech, licence Apache-2.0)

Spark-TTS (0.5B): Bilingual TTS (Text-to-Speech, licence Apache-2.0)

StyleTTS 2: Style-based TTS (Text-to-Speech, licence MIT + conditions)

Bark (Suno): Generative TTS (Text-to-Speech, licence MIT)

Parler-TTS: Instruction-following TTS (Text-to-Speech, licence Apache-2.0)

OuteTTS: Pure-LLM TTS (Text-to-Speech, licence Apache-2.0)

VibeVoice (Microsoft): Long-form expressive TTS (Text-to-Speech, licence Verify (open weights on GitHub))

GPT-SoVITS: Voice cloning TTS (Text-to-Speech, licence MIT)

OpenVoice (MyShell): Voice cloning (Text-to-Speech, licence MIT)

Sarvam (Saaras STT / TTS): Indic speech-to-text and text-to-speech models (Indic Stack, licence Commercial API (some open models))

Moshi (Kyutai): Full-duplex spoken dialogue (Speech-to-Speech, licence Weights CC-BY-4.0; code MIT/Apache)

PersonaPlex (NVIDIA): Moshi-based, persona-controllable (Speech-to-Speech, licence Weights NVIDIA Open Model License; code MIT)

Nemotron VoiceChat 11B (NVIDIA): Full-duplex + tool calls (Speech-to-Speech, licence OpenMDW 1.1)

Hertz-dev: Full-duplex dialogue (Speech-to-Speech, licence Apache-2.0)

dGSLM (Meta): Dual-tower dialogue (Speech-to-Speech, licence MIT)

Hibiki / Hibiki-Zero (Kyutai): Streaming speech translation (Speech-to-Speech, licence CC-BY-4.0 / MIT)

LiveKit Agents: WebRTC-first agent framework (Python/JS/Go) (Agent Frameworks, licence Apache-2.0)

Pipecat (Daily): Transport-agnostic pipeline framework (Python) (Agent Frameworks, licence BSD-2-Clause)

TEN Framework (Agora): Multimodal graph-based framework (Agent Frameworks, licence Apache-2.0)

Vocode: Voice agent framework (Agent Frameworks, licence MIT)

Bolna: Telephony + WebSocket voice agents (Agent Frameworks, licence Open source (verify exact license))

Silero VAD: Voice activity detection (VAD, Noise & Diarization, licence MIT)

Pipecat SmartTurn: ML turn-taking prediction (VAD, Noise & Diarization, licence Check repo (Pipecat BSD-2-Clause family))

Deepgram Flux: Turn-taking (VAD, Noise & Diarization, licence Commercial (Deepgram))

Speechmatics (EndOfUtteranceMode.ADAPTIVE): Adaptive end-of-utterance (VAD, Noise & Diarization, licence Commercial)

RNNoise: Neural noise suppression (VAD, Noise & Diarization, licence BSD-3-Clause)

DeepFilterNet: Neural noise suppression (VAD, Noise & Diarization, licence MIT)

WebRTC NS / AEC: Browser-grade noise suppression + echo cancellation (VAD, Noise & Diarization, licence BSD-3-Clause)

pyannote.audio: Speaker diarization (VAD, Noise & Diarization, licence Code MIT; pretrained weights gated — commercial use needs paid license)

NeMo Sortformer: Streaming diarization (VAD, Noise & Diarization, licence Apache-2.0 (toolkit))

Krisp: Noise/voice isolation (VAD, Noise & Diarization, licence Commercial)

Mimi (Kyutai): Streaming neural audio codec (12.5 Hz, 1.1 kbps) (Codecs & Tools, licence Verify (Kyutai family))

HiFi-GAN / Vocos / BigVGAN: Neural vocoders (Codecs & Tools, licence MIT (generally))

Whisper large-v3-turbo: Distilled Whisper (Speech-to-Text, licence MIT)

distil-whisper: Distilled Whisper (Speech-to-Text, licence Apache-2.0)

Coval: Voice AI eval framework (Evals & Serving, licence Verify)

Pipecat eval repo: Voice agent eval harness (Evals & Serving, licence BSD-2-Clause family)

ai-coustics: Noise suppression / voice cleanup (VAD, Noise & Diarization, licence Commercial)

LiveKit RecorderIO: In-process call recording (LiveKit Agents v1.3+) (Codecs & Tools, licence Apache-2.0 (LiveKit))

UNMUTE (Agentic AI Foundation): Declarative voice-agent standard (Agent Frameworks, licence Open source)

Breeze TTS 2: (Text-to-Speech, licence Research / non-commercial weights (code Apache-2.0))

Fish Audio S2 Pro: (Text-to-Speech, licence Research / non-commercial)

Step Audio EditX: (Text-to-Speech, licence Apache-2.0)

Voxtral TTS: (Text-to-Speech, licence CC BY-NC 4.0)

Magpie-Multilingual 357M: (Text-to-Speech, licence NVIDIA Open Model)

Maya1: (Text-to-Speech, licence Apache-2.0)

Higgs Audio V3: (Text-to-Speech, licence Research / non-commercial)

MetaVoice-1B: (Text-to-Speech, licence Apache-2.0)

VoxCPM2: (Text-to-Speech, licence Apache-2.0)

GLM-TTS: (Text-to-Speech, licence MIT weights)

Meta Omnilingual ASR: (Speech-to-Text, licence Apache-2.0)

Cohere Speech v1: (Speech-to-Text, licence MIT)

Kyutai STT: (Speech-to-Text, licence open)

MiniCPM-o 4.5: (Speech-to-Speech, licence Apache-2.0)

Realtime-Venus: (Speech-to-Speech, licence Apache-2.0)

Qwen3-Omni-30B-A3B: (Speech-to-Speech, licence Apache-2.0)

Qwen2.5-Omni-7B / 3B: (Speech-to-Speech, licence Apache-2.0)

Step-Audio 2 mini: (Speech-to-Speech, licence Apache-2.0)

Kimi-Audio-7B-Instruct: (Speech-to-Speech, licence MIT)

Fun-Audio-Chat-8B: (Speech-to-Speech, licence Apache-2.0)

GLM-4-Voice-9B: (Speech-to-Speech, licence GLM-4-Voice license)

LFM2.5-Audio-1.5B: (Speech-to-Speech, licence LFM Open v1.0)

Ultravox v0.7: (Speech-to-Speech, licence MIT)

LLaMA-Omni / Omni2: (Speech-to-Speech, licence open)

Mini-Omni / SLAM-Omni / Freeze-Omni: (Speech-to-Speech, licence open)

Kyutai Unmute: (Agent Frameworks, licence Open)

FireRedAudio: (Speech-to-Speech, licence open)

LocalAI: Self-hosted, OpenAI-compatible API server for local models (Evals & Serving, licence MIT)

vLLM-Omni: Production GPU serving for new-generation TTS and omni models (Evals & Serving, licence Apache-2.0)

llama.cpp-omni: GGUF CPU inference for speech models such as VoxCPM2 (Evals & Serving, licence MIT)

Kokoro-FastAPI: Ready-made API server for Kokoro TTS (Evals & Serving, licence Apache-2.0)

Speaches: Local server compatible with the OpenAI Realtime API (Evals & Serving, licence MIT)

voice-agent-next: Open voice-agent runtime with a latency benchmark harness (Evals & Serving, licence Check repo)

TEN VAD: Voice activity detection from the TEN framework (VAD, Noise & Diarization, licence Check repo)

AI4Bharat IndicConformer: 600M multilingual Conformer ASR covering all 22 scheduled Indian languages (Indic Stack, licence MIT)

Vani: Open protocol and Python SDK for Indian-language voice agents (Indic Stack, licence Apache-2.0)

VoicEra: Open voice-AI building block with telephony (Indic Stack, licence Check repo)

Bhashini VoicERA: India's national open-source voice AI stack (Feb 2026) (Indic Stack, licence Check license)

Gnani.ai Vachana: Indic speech-to-text trained on 1M hours under the IndiaAI Mission (Indic Stack, licence Check license)

Navana.ai RESPIN: 10,000-hour Indian speech dataset (Indic Stack, licence Check license)

Shunya Labs: Sovereign speech models covering 216 languages (Indic Stack, licence Check license)

SraVaani 1.0 / DhVaani 0.5: Indic speech models (Indic Stack, licence Check license)

Sooktam 2 / Shrutam 2: Indic speech models (Indic Stack, licence Check license)

Svara TTS v1: Open multilingual TTS for 19 languages: 18 Indian languages plus Indian English (Text-to-Speech, licence Apache-2.0)

How we built this list

We combined hundreds of real discussions among voice AI builders in the Unio community with 2026 benchmarks, including the Artificial Analysis Speech Arena and the Open ASR Leaderboard, and each project’s own repository.

Licence labels: OK to ship means a permissive licence such as MIT, Apache-2.0 or CC-BY-4.0. Check licence means custom, copyleft (GPL) or gated weights. Non-commercial means research-only weights. Licences change often, so always confirm in the repository before you ship.

Community picks are ranked by how often builders recommend a project and how positively they talk about it. Arena Elo scores come from blind listener votes; higher is better.

Using a model that’s missing, or found a licence change? Tell the Unio team and we’ll update the list. This page is updated every month.