New open speech-to-text models: autumn 2026 releases
7 open-weight speech-to-text models released or updated since our last database review are tracked here, from NVIDIA's 34B-parameter Nemotron 3.5 Transcribe to a 16.9 MB streaming model for microcontrollers. Every fact is copied from the official model card; where a card is silent, we say so. 6 of them are now in the matcher with provisional ratings; the rest are tracked only.
| Model | Released | License | Streaming | Size | Status |
|---|---|---|---|---|---|
| NVIDIA Nemotron 3.5 Transcribe 30B-A4B | October 28, 2026 | OpenMDW-1.1 | No (offline) | 34B total, 4.3B active, including the audio encoder | Tracked only |
| IBM Granite Speech 5.0 2B | October 20, 2026 | Apache-2.0 | No (offline) | 2B parameters | In matcher |
| IBM Granite Speech 5.0 470M TurboCTC | October 21, 2026 | Apache-2.0 | No (offline) | 470M parameters | In matcher |
| Microsoft VibeVoice-ASR-Streaming-7B | September 2026 | MIT | Yes | The card lists 9B parameters | In matcher |
| Moondream Parakeet Redux | Not stated | CC BY 4.0 | No (offline) | 1.58-bit (ternary) version of Parakeet TDT 0.6B v3 | In matcher |
| Cactus Whistle | Not stated | Apache-2.0 | Yes | A single 16.9 MB file | In matcher |
| IFM Jais-ASR | Not stated | Apache-2.0 | No (offline) | 5.7B total, including a 0.9B audio encoder | In matcher |
Model by model
NVIDIA Nemotron 3.5 Transcribe 30B-A4B
A server-class offline model that adds diarization and translation. It is a much larger relative of the reviewed Nemotron 3.5 ASR Streaming 0.6B, not a laptop model. Related reviewed model: Nemotron 3.5 ASR Streaming 0.6B.
- Released
- October 28, 2026
- Size
- 34B total, 4.3B active, including the audio encoder
- License
- OpenMDW-1.1
- Languages
- 40 languages, in support tiers
- Streaming
- No. NVIDIA calls it an offline model; internal chunking does not make it real-time streaming.
- Speakers
- Speaker-attributed transcription
- Timestamps
- Optional word timestamps
- Translation
- Yes, to and from English for many listed languages
- How to run
- vLLM 0.32.0 with the NeMo Speech main branch; Transformers installed from source
- Hardware
- Reference deployment: one H100 80 GB GPU
Tracked only, not scored in the matcher · Official model card ↗ · checked November 3, 2026
IBM Granite Speech 5.0 2B
A permissively licensed multilingual option with an unusually wide set of documented runtimes, including llama.cpp and Apple's MLX.
- Released
- October 20, 2026
- Size
- 2B parameters
- License
- Apache-2.0
- Languages
- English, French, German, Spanish, Portuguese, Japanese
- Streaming
- Not stated on the model card
- Speakers
- Not stated on the model card
- Timestamps
- Not stated on the model card
- Translation
- To and from English for the same languages, plus English to Italian and English to Mandarin
- How to run
- Transformers 4.52.1 or later, vLLM, llama.cpp (GGUF) and mlx-audio 0.5.9 or later
- Hardware
- No memory figure published. Transcribes segments of up to about 2 minutes; IBM advises splitting longer audio, for example with VAD.
Now in the matcher with provisional ratings: full profile · Official model card ↗ · checked November 3, 2026
IBM Granite Speech 5.0 470M TurboCTC
A compact, permissively licensed candidate for on-device transcription in seven languages.
- Released
- October 21, 2026 (an English-only first version appeared on August 25, 2026)
- Size
- 470M parameters
- License
- Apache-2.0
- Languages
- English, French, German, Spanish, Portuguese, Italian, Japanese
- Streaming
- Not stated on the model card
- Speakers
- Not stated on the model card
- Timestamps
- Not stated on the model card
- Translation
- Not stated on the model card
- How to run
- Native in Transformers 5.16.0 or later; mlx-audio 0.5.1 or later on Apple Silicon
- Hardware
- Non-autoregressive CTC model that IBM describes as suited to laptops, smartphones and other edge devices
Now in the matcher with provisional ratings: full profile · Official model card ↗ · checked November 3, 2026
Microsoft VibeVoice-ASR-Streaming-7B
The larger sibling of the reviewed VibeVoice-ASR Streaming 1.5B, with the same ten languages and customized hotwords. Related reviewed model: VibeVoice-ASR Streaming 1.5B.
- Released
- September 2026 (paper and collection dated September 2)
- Size
- The card lists 9B parameters
- License
- MIT
- Languages
- Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish
- Streaming
- Yes. It continuously transcribes who said what as speech arrives.
- Speakers
- Speaker-attributed (who said what)
- Timestamps
- Not stated on the model card
- Translation
- Not stated on the model card
- How to run
- Transformers (pip install -U transformers accelerate); full instructions in the microsoft/VibeVoice repository
- Hardware
- No memory figure published
Now in the matcher with provisional ratings: full profile · Official model card ↗ · checked November 3, 2026
Moondream Parakeet Redux
A CPU-friendly route to Parakeet's 25 languages, which NVIDIA otherwise documents only for its GPUs. Related reviewed models: Parakeet TDT 0.6B v3, Orukeet.
- Released
- Not stated on the model card
- Size
- 1.58-bit (ternary) version of Parakeet TDT 0.6B v3; 178 MB
- License
- CC BY 4.0, same as the original
- Languages
- The same 25 European languages as Parakeet TDT 0.6B v3
- Streaming
- Not stated on the model card
- Speakers
- Not stated on the model card
- Timestamps
- Not stated on the model card
- Translation
- Not stated on the model card
- How to run
- Photon runtime via pip install moondream (2.4.0 or later), on CPU, Apple GPU (MPS) or CUDA
- Hardware
- Vendor-reported 113× real time on eight x86 CPU cores (AMD EPYC 9575F), and about 38× on the CPU of an M2 MacBook Air
Now in the matcher with provisional ratings: full profile · Official model card ↗ · checked November 3, 2026
Cactus Whistle
A tiny multilingual streaming model for devices far smaller than a Raspberry Pi; unlike Moonshine's reviewed models, it is not English-only. Related reviewed model: Moonshine Tiny Streaming (English).
- Released
- Not stated on the model card (2026)
- Size
- A single 16.9 MB file; parameter count not stated
- License
- Apache-2.0
- Languages
- English, German, French, Spanish, Italian, Dutch, Polish
- Streaming
- Yes, live audio streams of any length
- Speakers
- Not stated on the model card
- Timestamps
- Word-level start, end and probability
- Translation
- Not stated on the model card
- How to run
- pip install cactus-needle; runs on CPU with no dependencies and no GPU
- Hardware
- Aimed at phones, wearables, robots, smart-home devices, cars and microcontrollers
Now in the matcher with provisional ratings: full profile · Official model card ↗ · checked November 3, 2026
IFM Jais-ASR
An Arabic dialect specialist. Among the reviewed models, Arabic coverage comes from Whisper and other general multilingual families.
- Released
- Not stated on the model card (2026)
- Size
- 5.7B total, including a 0.9B audio encoder; built on Jais-3-4B
- License
- Apache-2.0
- Languages
- Modern Standard Arabic and dialects from 19 Arab countries, with code-switching into English, French, Spanish and Italian
- Streaming
- Not stated on the model card
- Speakers
- Not stated on the model card
- Timestamps
- Not stated on the model card
- Translation
- No; the card says it is built to transcribe rather than translate
- How to run
- Transformers 5.16.1, or vLLM 0.23.0 with the bundled plugin (OpenAI-compatible transcription endpoint)
- Hardware
- Plan on roughly 12 GB of GPU memory in BF16; FP8 (7.68 GB) and 4-bit AWQ (5.85 GB) downloads are also published
Now in the matcher with provisional ratings: full profile · Official model card ↗ · checked November 3, 2026
Announced, but not open weights
- Mistral Voxtral 3 transcription models. Reported in October 2026. We found no open weights; Voxtral Mini 4B Realtime 2602 remains the open-weight Mistral model in our database.
- TII Falcon-ASR. Shown as a demo; we found no published weights or license.
- Qwen-Audio-3.1-ASR. We found a technical report but no open weights.
How new models reach the matcher
A model enters the matcher only after its explicit language list, license, streaming, timestamp and diarization claims and documented runtimes are verified against primary sources and it receives conservative hardware and workload ratings (see the methodology). Until then, compare a new release against the reviewed models with the benchmark tool on your own audio.
Common questions
What are the newest open speech-to-text models?
Recent open-weight releases we have checked include NVIDIA Nemotron 3.5 Transcribe 30B-A4B, IBM Granite Speech 5.0 2B, IBM Granite Speech 5.0 470M TurboCTC and Microsoft VibeVoice-ASR-Streaming-7B. Each entry above links to its official model card and lists only what that card states.
Are these models in the matcher?
6 of them are, since November 3, 2026: Granite Speech 5.0 2B, Granite Speech 5.0 470M TurboCTC, VibeVoice-ASR-Streaming-7B, Parakeet Redux, Whistle and Jais-ASR. Their languages, licenses and features are verified, but their hardware and workload ratings are provisional until independent benchmarks exist, and the matcher says so in every result. NVIDIA Nemotron 3.5 Transcribe 30B-A4B is tracked only.
Is there a new model for Arabic dialects?
Yes. IFM Jais-ASR covers Modern Standard Arabic and dialects from 19 Arab countries under Apache-2.0, and needs roughly 12 GB of GPU memory in BF16. See also the Arabic language guide for reviewed options.