How to choose an open-source speech-to-text model
Compare open speech-to-text models by languages, live streaming, speaker labels, memory and license: Whisper large-v3-turbo, Qwen3-ASR, Voxtral Realtime, VibeVoice-ASR and more.
Short answer. If you want a safe default for files on a laptop, use Whisper large-v3-turbo with whisper.cpp. Choose Voxtral Realtime or Nemotron 3.5 ASR for live captions, Qwen3-ASR for Chinese dialects and singing, Jais-ASR for Arabic and its dialects, and VibeVoice-ASR or Nemotron 3.5 Transcribe when you need speaker labels. Whistle fits in 16.9 MB for a phone or a CPU-only device. Accuracy claims are the developers’ own, so test on your audio.
What a speech-to-text model does
Speech recognition turns audio into text. Open models differ in what comes with the transcript as well as in accuracy. Ask four questions before you compare numbers:
- Offline or live? A finished recording lets a model see all the audio, while live captions must appear while someone is speaking.
- How long is the audio? Some models are built for an hour in one pass, and others for short clips.
- Do you need speaker labels, timestamps or translation?
- Which languages and accents? Coverage runs from 13 to 50+ languages, and quality differs inside each list.
Our speech recognition models list shows every model we track.
The shortlist
| Model | Size | License | Languages | What stands out |
|---|---|---|---|---|
| Whisper large-v3-turbo | 0.8B | MIT | Many | About 8x faster than large; runs in roughly 6 GB of GPU memory |
| Qwen3-ASR | 0.6B, 1.7B | Apache 2.0 | 30 plus 22 Chinese dialects | Singing, streaming, word timestamps with the aligner |
| Voxtral Realtime | 4B | Apache 2.0 | 13 | Delay you set between 80 ms and 2.4 s |
| VibeVoice-ASR | 8.7B | MIT | 50+ | An hour of audio in one pass, speaker labels, hotwords |
| Nemotron 3.5 Transcribe | 30B-A4B | OpenMDW | 40 | Speaker labels, translation, prompt-controlled style; offline only |
| Nemotron 3.5 ASR | 0.6B | OpenMDW | 40 language-locales, 32 ready to use | Streaming, with the delay chosen per request from 80 ms to 1.12 s |
| Parakeet TDT 0.6B v3 | 0.6B | CC BY 4.0 | 25 European | Detects the language by itself; word timestamps; credit NVIDIA |
| Cohere Transcribe | 2B | Apache 2.0 | 14 | Transcription only; you name the language; gated download |
| Jais-ASR | 5.7B | Apache 2.0 | Arabic with dialects from 19 countries, plus English, French, Spanish and Italian | Code-switching, optional diacritics, dialect labels; best average on an Arabic leaderboard in IFM’s own table |
| Granite Speech 5.0 | 2B, 470M | Apache 2.0 | 6 and 7 | The 2B translates to and from English; the 470M CTC model is built for speed |
| MOSS-Transcribe-Diarize | 0.9B | Apache 2.0 | 50+ | Up to 90 minutes in one pass, speaker labels, hotwords |
| GigaAM-v3 | 220M to 240M | MIT | Russian | Punctuated, normalized output in the e2e versions |
| GigaAM Multilingual | 220M, 600M | MIT | Reported for English, Russian, Kazakh, Kyrgyz, Uzbek | Strong on Central Asian languages in Sber’s tests; English is only moderate |
| Whistle | 55M | Apache 2.0 | 7 | A 16.9 MB file that runs on a plain CPU; word timestamps |
| LFM3-Audio | 2B | LFM Open License | 18 | Speech in and out on-device; transcription and translation |
Pick by the job
- Meetings, interviews and podcasts with several speakers. VibeVoice-ASR takes up to an hour in one pass and adds speaker labels and timestamps. Nemotron 3.5 Transcribe also labels speakers, but only in 13 of its 40 languages. MOSS-Transcribe-Diarize is a 0.9B Apache 2.0 alternative that handles up to 90 minutes, though its published scores come from Chinese-language sets. See speaker diarization for what the labels mean.
- Live captions, voice assistants and call agents. Voxtral Realtime lets you trade delay for accuracy. Nemotron 3.5 ASR does the same in a 0.6B model, with 32 of its 40 language-locales ready to use. Qwen3-ASR has a streaming mode, and VibeVoice-ASR ships streaming versions for 10 languages.
- European languages on a small GPU. Parakeet TDT 0.6B v3 covers 25 European languages, detects which one is spoken and gives word timestamps. Parakeet TDT 0.6B v3 vs Cohere Transcribe sets it beside the 14-language transcription specialist.
- Russian and Central Asian languages. GigaAM-v3 is a Russian recognizer whose e2e versions write punctuation, and GigaAM Multilingual adds Kazakh, Kyrgyz and Uzbek. Sber’s own tests put them ahead of Whisper large v3, by a wide margin on Kazakh, Kyrgyz and Uzbek, but all scores are the developer’s and the Multilingual model’s English is only moderate.
- A phone, a wearable or a CPU-only device. Whistle is a 16.9 MB model for seven European languages, and Granite Speech 5.0 470M TurboCTC is built for speed on edge devices.
- Arabic and its dialects. Jais-ASR is built for Arabic as it is spoken: Modern Standard Arabic and dialects from 19 countries, with code-switching into English, French, Spanish or Italian, optional diacritics and a dialect label on each transcript. In IFM’s own table it has the lowest average error on the Open Universal Arabic ASR Leaderboard, 22.85% against 23.17% for the next open model, though it trails the best systems on clean read speech. It needs about 12 GB of GPU memory in BF16, and a 5.8 GB AWQ build exists.
- Chinese, dialects and song lyrics. Qwen3-ASR covers 22 Chinese dialects and transcribes singing, and it offers a forced aligner for word or character timestamps.
- A laptop or an offline subtitle workflow. Whisper large-v3-turbo gives sentence and word timestamps, and many apps and runtimes support it. The turbo model does not translate, so use large-v3 for speech translation into English.
- Medical, legal or technical vocabulary. VibeVoice-ASR accepts hotwords such as names and terms, and Nemotron 3.5 Transcribe accepts the terms you expect.
- Translation. Nemotron 3.5 Transcribe translates 31 of its 40 languages, Granite Speech 5.0 2B translates to and from English, and LFM3-Audio translates among its 18 languages on-device.
Compare latency and accuracy together
For live use, delay and error rate move in opposite directions. Voxtral Realtime’s own table shows an average error rate of 12.60% at a 160 ms delay and 8.72% at 480 ms, so you can choose the point that suits your product. The same table shows how much language matters: at 480 ms, English scores 4.90%, Korean 15.74% and Arabic 22.53%. Qwen3-ASR vs Voxtral Realtime puts the two live-capable models side by side.
Word error rate is the standard score, and the glossary entry word error rate explains how it is computed and why the same model can score very differently on clean read speech and on real meetings.
Check memory and speed
In BF16, weights take about 2 bytes per parameter, and running the model needs more for activations:
| Model | Parameters | Weights in BF16 | Note |
|---|---|---|---|
| Qwen3-ASR 0.6B | 0.94B | about 1.9 GB | The fast option for batch jobs |
| Whisper large-v3-turbo | 0.81B | about 1.6 GB | OpenAI’s README lists roughly 6 GB of GPU memory to run it |
| Qwen3-ASR 1.7B | 2.35B | about 4.7 GB | |
| Voxtral Realtime | 4.43B | about 8.9 GB | |
| VibeVoice-ASR | 8.67B | about 17 GB | Microsoft also offers a CPU engine |
| Granite Speech 5.0 470M TurboCTC | 0.47B | about 0.95 GB | The small, fast CTC model |
| Nemotron 3.5 ASR | 0.64B | about 1.3 GB | The Hugging Face checkpoint is stored in FP32, at 2.55 GB |
| Parakeet TDT 0.6B v3 | 0.63B | about 1.3 GB | The checkpoint is FP32, at 2.51 GB; the 8-bit GGUF is 0.71 GB |
| Cohere Transcribe | 2.07B | about 4.1 GB | |
| Jais-ASR | 5.72B | about 11.4 GB | The card plans on roughly 12 GB of GPU memory; a 5.8 GB AWQ build exists |
| Nemotron 3.5 Transcribe | 34B | about 68 GB | About 22 GB at NVFP4; the card targets a single H100 80 GB |
Use the VRAM calculator for your own setup.
Know the limits
Recognizers make characteristic mistakes. Whisper can invent words that were never spoken, especially in silence or noise, and any model can mishear names and numbers. Accuracy drops with accents, overlapping speech and poor microphones. Run a sample of your own recordings, listen to the failures, and check the proper nouns your users care about. How to evaluate AI models yourself describes how to build a small test set.
Run it
whisper.cpp runs Whisper models on CPUs, Apple Silicon and phones. Newer models ship with their own runtimes, so read the model card first: Voxtral Realtime’s card recommends vLLM with websocket streaming, which needs more setup than a simple file-transcription tool. Parakeet TDT 0.6B v3 and Nemotron 3.5 ASR run in NVIDIA’s NeMo toolkit, and Whistle comes with its developer’s own C++ engine. If you also want your application to talk back, how to choose a text-to-speech model covers the other half of a voice pipeline.
Frequently asked questions
- What is the best open-source speech-to-text model?
- It depends on your audio. Whisper large-v3-turbo is the most downloaded, with about 5.1 million downloads in the 30 days to 11 November 2026, and the best supported by apps and runtimes. Newer models add things Whisper does not offer natively, such as live streaming, hour-long input and speaker labels, but all of their accuracy claims come from the developers.
- Which open models can transcribe live audio?
- Voxtral Realtime lets you set the delay between 80 ms and 2.4 s, and Nemotron 3.5 ASR between 80 ms and 1.12 s. Qwen3-ASR has a streaming mode, and VibeVoice-ASR has streaming versions that cover 10 languages. Nemotron 3.5 Transcribe is offline only, so it needs finished recordings.
- Which models label who is speaking?
- VibeVoice-ASR, MOSS-Transcribe-Diarize and Nemotron 3.5 Transcribe return speaker labels, and the Nemotron card limits speaker attribution to 13 languages. The Qwen3-ASR and Voxtral Realtime cards do not describe speaker labels, so use a separate diarization tool with them.
- How accurate are open speech-to-text models?
- Accuracy varies by language, accent and noise, and every figure on a model card comes from the developer. Voxtral Realtime's own table shows 4.90% word error on English and 22.53% on Arabic at a 480 ms delay. Test on your own audio, and check names and numbers by hand.
- Can I use these models in a commercial product?
- Yes for most. Whisper, VibeVoice-ASR, GigaAM-v3 and GigaAM Multilingual are MIT, Qwen3-ASR, Voxtral Realtime, Cohere Transcribe, Jais-ASR, Granite Speech 5.0, MOSS-Transcribe-Diarize and Whistle are Apache 2.0, and the two Nemotron models use the permissive OpenMDW license. Parakeet TDT 0.6B v3 is CC BY 4.0, so you must credit NVIDIA. LFM3-Audio is free to use commercially only below US$10 million of yearly revenue.
- Which open speech-to-text models run on a CPU or a phone?
- Whistle is a 16.9 MB model for seven languages that its developer runs on a plain CPU, reaching the first token in about 11 ms on an Apple M4 Pro. IBM built Granite Speech 5.0 470M TurboCTC for speed and edge devices, and its BF16 weights are under 1 GB. Parakeet TDT 0.6B v3 also has an 8-bit GGUF file of 0.71 GB for a lightweight C++ runtime.
Key terms in this guide
- Benchmark A benchmark is a standardized test used to compare models, such as a set of maths, coding or knowledge questions; scores are useful but easy to over-read.
- On-device AI On-device AI means the model runs on your own phone, laptop or small board instead of in the cloud, so data stays local and it works offline.
- Speaker diarization (speaker labels) Diarization works out who spoke when in a recording and labels each stretch of speech with a speaker, such as Speaker 1 and Speaker 2.
- Speech recognition (ASR, speech-to-text) Speech recognition, also called ASR or speech-to-text, turns spoken audio into written text.
- Word error rate (WER) Word error rate is the standard score for speech recognition: the share of words in a transcript that are wrong, missing or extra compared with a correct reference.
Models mentioned in this guide
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Whisper large-v3-turbo OpenAI | 810M – 1.54B | – | MIT (Permissive) | 6 GB Laptop or phone | |
| Qwen3-ASR Alibaba Qwen | 918M – 2.35B | – | Apache 2.0 (Permissive) | 2.12 GB Laptop or phone | |
| Voxtral Realtime Mistral AI | 4.43B | 128K | Apache 2.0 (Permissive) | 10 GB Laptop or phone | |
| VibeVoice-ASR Microsoft | 2.81B – 8.67B | 64K | MIT (Permissive) | 6.47 GB Laptop or phone | |
| Nemotron 3.5 Transcribe NVIDIA | 34B 4.3B active | 64K | OpenMDW-1.1 (Permissive) | 25 GB Gaming GPU | |
| Nemotron 3.5 ASR NVIDIA | 638M | – | OpenMDW-1.1 (Permissive) | 2.93 GB Laptop or phone | |
| Parakeet TDT 0.6B v3 NVIDIA | 627M | – | CC BY 4.0 (Permissive) | 2.89 GB Laptop or phone | |
| Cohere Transcribe Cohere | 2.07B | – | Apache 2.0 (Permissive) | 4.75 GB Laptop or phone | |
| Jais-ASR Institute of Foundation Models (MBZUAI) | 5.72B | – | Apache 2.0 (Permissive) | 3.63 GB Laptop or phone | |
| Granite Speech 5.0 IBM | 470M – 2.31B | – | Apache 2.0 (Permissive) | 1.09 GB Laptop or phone | |
| Whistle Cactus Compute | 55M | – | Apache 2.0 (Permissive) | 0.253 GB Laptop or phone | |
| MOSS-Transcribe-Diarize OpenMOSS | 910M | 128K | Apache 2.0 (Permissive) | 2.09 GB Laptop or phone | |
| GigaAM-v3 Sber (GigaChat) | 220M – 240M | – | MIT (Permissive) | 0.506 GB Laptop or phone | |
| GigaAM Multilingual Sber (GigaChat) | 220M – 600M | – | MIT (Permissive) | 1.01 GB Laptop or phone | |
| LFM3-Audio-2B Liquid AI | 2B | 32K | LFM Open 1.0 (Conditional) | 4.6 GB Laptop or phone |