Open-source speech models,
can you actually run them?
Can I run it, on what hardware, and can I use it commercially? 38 TTS and STT models with license, memory, streaming and cloning, every fact linked to its source. Speed is measured by us, not copied from the press release.
38models
22text-to-speech
16speech-to-text
31commercial OK
2benchmarked
| Model | Type | Released | Params | Commercial | Min memory | Stream | Clone | Languages | Speed (RTF) | Cmp |
|---|---|---|---|---|---|---|---|---|---|---|
| Fish Audio Drama 3 Fish Audio |
TTS | 2026-10-01 | 35B MoE (3B active per token) | Conditional | – | Yes | Yes | 83 languages | pending | |
| Higgs TTS 3 4B Boson AI |
TTS | 2026-06-04 | 4.65B | Conditional | – | Yes | Yes | 100+ languages | pending | |
| Miso TTS 8B Miso Labs |
TTS | 2026-05-21 | ~8.2B | Yes | 24 GB VRAM recommended (bf16); ~20 GB RAM on CPU (slow) | Unknown | Yes | English | pending | |
| Supertonic 3 Supertone |
TTS | 2026-04-29 | ~99M | Yes | – CPU | Unknown | No | 31 languages | pending | |
| Parakeet Unified EN 0.6B NVIDIA |
STT | 2026-04-07 | 600M | Yes | – | Yes | n/a | English | pending | |
| Fish Audio S2 Pro Fish Audio |
TTS | 2026-03-09 | 4B slow AR + 400M fast AR | No | – | Yes | Yes | 83 languages | pending | |
| KittenTTS Mini 0.8 KittenML |
TTS | 2026-02-19 | 80M | Yes | – CPU | Unknown | Unknown | – | pending | |
| Qwen3-ASR 1.7B Alibaba Qwen |
STT | 2026-01-28 | 2.35B | Yes | – | Yes | n/a | 30 languages + 22 Chinese dialects (52 total) | pending | |
| LuxTTS YatharthS |
TTS | 2026-01-22 | – | Yes | Fits within 1 GB VRAM CPU | Unknown | Yes | English | pending | |
| Voxtral Mini 4B Realtime (2602) Mistral AI |
STT | 2026-01-21 | 4.43B | Yes | Single GPU with >= 16 GB memory (BF16, vLLM) | Yes | n/a | 13 languages | pending | |
| VibeVoice-ASR Microsoft |
STT | 2026-01-21 | 8.67B | Yes | – | No | n/a | 50+ languages | pending | |
| Qwen3-TTS 1.7B Base Alibaba Qwen |
TTS | 2026-01-21 | 1.93B | Yes | – | Yes | Yes | 10 languages | pending | |
| Moonshine Streaming Medium Moonshine AI (Useful Sensors) |
STT | 2026-01-06 | 245M | Yes | Targets sub-1 GB memory edge devices CPU | Yes | n/a | English | pending | |
| Pocket TTS Kyutai |
TTS | 2025-12-29 | 100M | Yes | – CPU | Yes | Yes | 7 languages | pending | |
| Fun-CosyVoice3 0.5B (2512) Alibaba FunAudioLLM |
TTS | 2025-12-11 | 0.5B | Yes | – | Yes | Yes | 9 languages | pending | |
| VibeVoice-Realtime 0.5B Microsoft |
TTS | 2025-12-04 | 0.5B LLM (1.02B total in safetensors) | Yes | – | Yes | Unknown | English (+ experimental voices in 9 more languages) | pending | |
| Chatterbox-Turbo Resemble AI |
TTS | 2025-12-02 | 350M | Yes | – | Unknown | Yes | English | pending | |
| NeuTTS Air Neuphonic |
TTS | 2025-09-15 | ~552M (emb + active), ~360M active | Yes | – CPU | Yes | Yes | English | pending | |
| Parakeet TDT 0.6B v3 NVIDIA |
STT | 2025-08-04 | 600M | Yes | At least 2 GB RAM to load | Yes | n/a | 25 languages | pending | |
| Canary 1B v2 NVIDIA |
STT | 2025-08-04 | 978M | Yes | At least 6 GB RAM to load | Unknown | n/a | 25 languages | pending | |
| Kyutai TTS 1.6B (EN/FR) Kyutai |
TTS | 2025-06-30 | 1.8B | Yes | – | Yes | Limited | English, French | pending | |
| Canary-Qwen 2.5B NVIDIA |
STT | 2025-06-26 | 2.56B | Yes | – | Unknown | n/a | English | pending | |
| IndexTTS-2 bilibili (IndexTeam) |
TTS | 2025-06-18 | – | Conditional | – | Unknown | Yes | Chinese, English | pending | |
| Kyutai STT 2.6B (EN) Kyutai |
STT | 2025-06-06 | 2.62B | Yes | – | Yes | n/a | English | pending | |
| Chatterbox (Multilingual) Resemble AI |
TTS | 2025-04-24 | 500M | Yes | – | Unknown | Yes | 23+ languages | pending | |
| Granite Speech 3.3 8B IBM |
STT | 2025-04-14 | 8.65B | Yes | – | Unknown | n/a | 6 languages | pending | |
| Orpheus 3B (0.1 ft) Canopy Labs |
TTS | 2025-03-17 | 3.78B | Yes | – | Yes | Yes | English | pending | |
| Sesame CSM 1B Sesame |
TTS | 2025-03-06 | 1.55B | Yes | – | Unknown | Unknown | English | pending | |
| Kokoro 82M hexgrad |
TTS | 2024-12-26 | 82M | Yes | – | Unknown | Unknown | 9 languages | pending | |
| Distil-Whisper large-v3.5 Hugging Face |
STT | 2024-12-05 | 756M | Yes | – | Unknown | n/a | English | pending | |
| Moonshine Base Moonshine AI (Useful Sensors) |
STT | 2024-11-02 | 61M | Yes | – CPU | No | n/a | English | pending | |
| F5-TTS SWivid (SJTU et al.) |
TTS | 2024-10-07 | – | No | – | Unknown | Yes | – | pending | |
| Whisper large-v3-turbo OpenAI |
STT | 2024-10-01 | 809M | Yes | ~6 GB VRAM | No | n/a | 99 languages | pending | |
| faster-whisper (large-v3, CTranslate2) SYSTRAN |
STT | 2023-11-23 | 1.55B | Yes | ~2.9 GB VRAM (large-v2, int8, beam 5, project benchmark) CPU | No | n/a | 100 languages | pending | |
| Whisper large-v3 OpenAI |
STT | 2023-11-07 | 1.54B | Yes | ~10 GB VRAM | No | n/a | 99 languages | pending | |
| XTTS v2 Coqui |
TTS | 2023-10-31 | – | No | – | Yes | Yes | 17 languages | pending | |
| Piper (voices) Rhasspy / Open Home Foundation |
TTS | 2023-06-22 | – | Conditional | – CPU | Unknown | Unknown | 35 languages | 0.017 CPU |
|
| Whisper tiny OpenAI |
STT | 2022-09-26 | 38M | Yes | ~1 GB VRAM CPU | No | n/a | 99 languages | 0.030 CPU |
RTF = processing time ÷ audio length (lower is faster; <1 = faster than real time). "Unknown" means the model card or repo doesn't state it. Commercial: conditional = allowed only under conditions (revenue cap, GPL, contact vendor…), see each model page.