Best open-source speech models for transcription and text-to-speech
Speech models turn audio into text and text into audio, and most of them are small enough for a laptop. This page covers both directions. Our dedicated lists for text-to-speech and speech recognition go deeper on each direction. Always spot-check transcripts and be careful with voice cloning: using someone's voice without permission can be illegal.
-
Nemotron 3.5 Transcribe
NVIDIA's speech-to-text model with 40 languages, speaker labels, translation and prompt-controlled transcript style, under the permissive OpenMDW license. It processes finished recordings rather than live streams.
-
Whisper large-v3-turbo
OpenAI's MIT-licensed speech recognition model. The turbo version is about 8x faster than large and needs roughly 6 GB of GPU memory. It can hallucinate text that was never spoken, so check important transcripts.
-
Qwen3-ASR
Alibaba's Apache 2.0 speech-to-text pair (1.7B and 0.6B): 30 languages plus 22 Chinese dialects, singing, streaming and timestamps.
-
Cohere Transcribe
Cohere's Apache 2.0 2B model that only transcribes: 14 languages, with no timestamps, no speaker labels and no automatic language detection.
-
Voxtral Realtime
Mistral's Apache 2.0 4B streaming speech-to-text model: 13 languages and a delay you set between 80 ms and 2.4 s, for live use.
-
Confucius4-R2T2
NetEase Youdao's streaming speech-to-text model for Chinese and English, adapted from Qwen3-ASR-1.7B. It never revises committed words, and the license needs a separate agreement above 100 million monthly users or RMB 1 billion of revenue.
-
Qwen3-TTS
Alibaba's Apache 2.0 text-to-speech family in 1.7B and 0.6B sizes: 10 languages, voice cloning from about 3 seconds of audio, voice design from a description, and streaming latency as low as 97 ms.
-
Kokoro Most downloaded
hexgrad's Apache 2.0 82M text-to-speech model with 54 preset voices in 8 languages. It is the most downloaded speech model here, with about 12.5 million downloads in the 30 days to 11 November 2026, but it cannot clone voices.
-
VoxCPM2
OpenBMB's 2B Apache 2.0 text-to-speech model: 30 languages, voice design from a text description, cloning and 48 kHz output on about 8 GB of VRAM.
-
IndexTTS-2.5
Bilibili's 0.8B zero-shot cloner for Chinese, English, Japanese, Spanish and Arabic, with separate emotion, pronunciation and speed controls on about 6 GB of VRAM. Its own license has no general commercial ban but sets 100 million-user and RMB 1 billion thresholds.
-
AuK
Tencent's MIT-licensed 1.5B model that handles speech generation, editing, enhancement and separation as written instructions. Peak memory is about 25 GiB in bf16, or about 17 GiB with CPU offload.
-
ZONOS3-Mini
Zyphra's Apache 2.0 800M text-to-speech model with voice cloning in 43 languages, released on 9 November 2026. Zyphra says it runs in real time on laptops, phones and in a browser.
-
Pocket TTS
Kyutai's 100M text-to-speech model with voice cloning that runs faster than real time on a laptop CPU. The weights are CC BY 4.0 behind a prohibited-use gate.
-
Drama 3
Fish Audio's 36B text-to-speech mixture of experts with 83 languages and performance directions written into the text. The license is free only below US$500,000 of yearly revenue, and it never covers a hosted voice service.
-
OmniVoice
Xiaomi's text-to-speech model covers 600+ languages with cloning and voice design. The code is Apache 2.0, but the weights are CC BY-NC, so it is for non-commercial use.
-
LFM3-Audio-2B
Liquid AI's 2B on-device voice model: speech in and out, transcription, synthesis and translation across 18 languages. The license is free only below US$10 million of revenue.
-
Sophon-Realtime
Adventists AI's open full-duplex voice model, released on 8 November 2026 under Apache 2.0: a 32B thinker that listens while you talk, handles interruptions and can laugh or sigh, with three builds and 0.25 to 0.6 s to first audio. It needs one 80 GB GPU, and the benchmark figures are the developer's own.
How we chose
- Open weights for speech recognition, speech synthesis or both.
- License terms read in full, with voice-cloning and revenue conditions noted.
- Language counts and latency figures are the developers' own claims.