Local speech-to-text models: the open Whisper alternatives
Qwen3-ASR 1.7B has the lowest English word error rate of the open models on the Open ASR Leaderboard: 5.40% averaged over 11 English sets (version 25 October 2026), under Apache-2.0. Whisper large-v3-turbo is less accurate (7.64%) but runs in more local tools than any other model. With no GPU, start with Moonshine v2 or Fun-ASR-Nano.
Written by Jakub Rusinowski · Last updated
The models
The licence column is the licence on the weights, read from each model card — not the badge on the code repository. Memory is what the maker or a named source publishes, not a measurement and not a fit verdict.
9 of 9 models
| Model | Hardware tier | Licence (weights) | CPU only | Languages | Runtimes | Published memory | English WER (Open ASR Leaderboard, 25 October 2026) |
|---|---|---|---|---|---|---|---|
| Whisper large-v3-turbo OpenAI · 809M | 8 GB GPU / 16 GB Mac No GPU | C MIT Commercial use allowed | Yes | 90 | whisper.cpp faster-whisper WhisperKit | Not published | 7.64% RTFx 794 |
| Qwen3-ASR 1.7B Alibaba Qwen · 1.7B | 16–24 GB GPU / 32 GB+ Mac 8 GB GPU / 16 GB Mac | C Apache-2.0 Commercial use allowed | — | 52 | Transformers vLLM FunASR | Not published | 5.40% RTFx 819 |
| Qwen3-ASR 0.6B Alibaba Qwen · 0.6B | 8 GB GPU / 16 GB Mac No GPU | C Apache-2.0 Commercial use allowed | — | 52 | Transformers vLLM FunASR | Not published | 6.24% RTFx 745 |
| Cohere Transcribe Cohere Labs · 2B | 16–24 GB GPU / 32 GB+ Mac 8 GB GPU / 16 GB Mac | C Apache-2.0 Commercial use allowed | — | 14 | Transformers vLLM | Not published | 5.77% RTFx 902 |
| Canary-Qwen 2.5B NVIDIA · 2.5B | 16–24 GB GPU / 32 GB+ Mac | C CC-BY-4.0 — Attribution required. Commercial use allowed | — | 1 | NeMo | Not published | 5.94% RTFx 862 |
| Parakeet TDT v3 NVIDIA · 0.6B | 8 GB GPU / 16 GB Mac | C CC-BY-4.0 — Attribution required. Commercial use allowed | — | 25 | NeMo Transformers | ~2 GB (published by NVIDIA (model card)) At least 2 GB RAM to load (card). | 6.18% RTFx 6,132 |
| Moonshine v2 Moonshine AI | No GPU | C MIT Commercial use allowed | Yes | — | ONNX (Moonshine Voice) Transformers | Not published | — |
| Fun-ASR-Nano FunAudioLLM (Alibaba Tongyi) · Qwen3-0.6B decoder + encoder | No GPU 8 GB GPU / 16 GB Mac | C Apache-2.0 Commercial use allowed | Yes | 31 | FunASR llama.cpp runtime (single binary, no Python) vLLM | ~0.95 GB (published by nile2000 on Hugging Face) single source encoder f16 470 MB + Qwen3-0.6B q4_K_M 484 MB GGUF | 7.24% RTFx 496 |
| Granite Speech 5.0 IBM · 2B | 8 GB GPU / 16 GB Mac | C Apache-2.0 Commercial use allowed | — | 6 | Transformers vLLM | Not published | 5.68% RTFx 545 |
No model matches these filters.
How to run it
Qwen3-ASR as an OpenAI-compatible server
qwen-asr-serve Qwen/Qwen3-ASR-1.7B --gpu-memory-utilization 0.8 --host 0.0.0.0 --port 8000From the Qwen3-ASR model card. Needs a GPU. Qwen3-ASR card
Whisper through Transformers
pip install --upgrade transformers datasets[audio] accelerateFrom the Whisper large-v3-turbo card; whisper.cpp and faster-whisper are the faster local runtimes. Whisper card
Parakeet through NeMo
pip install -U nemo_toolkit['asr']From the Parakeet TDT 0.6B v3 card. Parakeet card
Sources: model cards on Hugging Face, the Open ASR Leaderboard, the Artificial Analysis Speech Arena and the Open TTS Leaderboard. Checked 25 October 2026.
Frequently asked questions
What is the best open speech-to-text model?
On the Open ASR Leaderboard (version 25 October 2026), Qwen3-ASR 1.7B has the lowest average English word error rate among open models, at 5.40%. Granite Speech 5.0 2B (5.68%) and Cohere Transcribe (5.77%) follow.
Is Whisper still worth using?
Yes, for its ecosystem: whisper.cpp, faster-whisper and WhisperKit cover almost every platform. It is no longer the most accurate: Whisper large-v3-turbo scores 7.64% on the same board.
Which speech-to-text model runs without a GPU?
Moonshine v2 is built for edge devices and streams; Fun-ASR-Nano ships as a single llama.cpp-style binary with no Python; Whisper through whisper.cpp also runs on a CPU. Parakeet TDT 0.6B v3 needs about 2 GB of RAM to load, per its card.
Can I use these models commercially?
Every speech-to-text row on this page has weights that allow commercial use. Canary-Qwen and Parakeet are CC-BY-4.0, which asks for attribution.
What does WER mean?
Word error rate: the share of words the model gets wrong against a reference transcript. Lower is better. The leaderboard averages it over 11 English sets, two of them private, so its numbers are not comparable with older 7-set figures.