What is the best local speech-to-text model in 2026?

There is no single winner as of October 2026; it depends on the box. For a laptop or server with a GPU and 13 target languages, Mistral's Voxtral Realtime (4B, Apache 2.0) gives native streaming and the lowest reported WER. For Apple silicon dictation across 25 European languages, Moondream's Parakeet Ultra (600M, September 22, 2026) is the accuracy pick. For English-only maximum accuracy, NVIDIA Canary-Qwen 2.5B leads the Open ASR Leaderboard. For 99-language coverage, Whisper large-v3-turbo via whisper.cpp remains the default.

Is Whisper still the best local STT model?

Not on accuracy. Whisper large-v3 (1.55B) has been passed on English WER by NVIDIA's Parakeet and Canary-Qwen models and on 13-language WER by Voxtral Transcribe 2, and it has no native streaming. It still wins on breadth — 99 languages, MIT licence, and the whisper.cpp / faster-whisper / WhisperX ecosystem — so it remains the safest choice when you need a language the newer models do not cover.

What is the smallest usable speech-to-text model?

Cactus Whistle, released October 2, 2026, at 16.9 MB. It runs on CPU with no dependencies, covers English, German, French, Spanish, Italian, Dutch and Polish, reaches the first token in about 11 ms on an M4 Pro, and beats the 145 MB Whisper Base on LibriSpeech and FLEURS. Moonshine Tiny v2 (41.9 MB) is the smallest English streaming model.

Which local STT model should I use for real-time captions?

Voxtral Realtime for multilingual streaming with configurable 80 ms to 2.4 s latency, NVIDIA Parakeet Unified 0.6B for English streaming at about 160 ms with punctuation, Parakeet Realtime EOU 120M when you need end-of-utterance detection at 80–160 ms and can live without punctuation, or Moonshine for constrained edge devices. Whisper is not natively streaming and needs chunking tooling.

Quick Answer

Best Local Speech-to-Text Models 2026: Top 7 Ranked

Published: October 4, 2026

The short answer

As of October 2026 the local speech-to-text field has split by workload: Voxtral Realtime for multilingual streaming, Parakeet Ultra for Apple-silicon dictation, Canary-Qwen 2.5B for English accuracy, Whisper large-v3-turbo for language breadth, and Whistle or Moonshine for devices measured in megabytes. Whisper is no longer the accuracy leader in any category, but it is still the only model on this list that covers 99 languages. Facts verified October 4, 2026.

The ranking

#	Model	Maker / date	Params / size	Languages	Streaming	Licence	Best for
1	Voxtral Realtime	Mistral, Feb 4, 2026	4B (~2.5 GB Q4)	13	✅ native, 80 ms–2.4 s	Apache 2.0	Multilingual live transcription, 3-hour audio
2	Parakeet Ultra	Moondream, Sep 22, 2026	600M	25 European	Chunked (VAD head)	NVIDIA/Moondream open	Apple-silicon and GPU dictation
3	Canary-Qwen 2.5B	NVIDIA, 2025	2.5B	English	❌	CC-BY-4.0	Maximum English accuracy
4	Whisper large-v3-turbo	OpenAI, Oct 2024	809M (~3 GB Q5)	99	❌ (chunked)	MIT	Language breadth, ecosystem
5	Parakeet Unified 0.6B	NVIDIA, Apr 2026	600M	English	✅ ~160 ms	NVIDIA open	English streaming with punctuation
6	Moonshine Tiny v2	Useful Sensors, 2025	41.9 MB	English	✅	MIT	Raspberry Pi-class edge
7	Whistle	Cactus, Oct 2, 2026	16.9 MB	7	❌ (30 s pass)	Open	Microcontrollers, wearables, voice-to-tool-calls

Why each one is ranked where it is

1. Voxtral Realtime (Mistral). Launched February 4, 2026 as half of Voxtral Transcribe 2 (the other half, Mini Transcribe V2, is API-only). It is the only open model here with a causal audio encoder trained for streaming, so latency is a dial from 80 ms to 2.4 s rather than a chunking hack, and it handles about three hours of audio natively through a 131K-token window. On FLEURS across its 13 languages it reports roughly 5.9% average WER against Whisper large-v3’s 7.4%. The costs: 16 GB VRAM in BF16 (about 2.5 GB quantised), and 13 languages is a hard ceiling.

2. Parakeet Ultra (Moondream). Moondream released Ultra and Redux on September 22, 2026 as post-trained variants of NVIDIA’s Parakeet TDT 0.6B V3, keeping V3’s architecture and tokenizer and adding a small voice-activity head so the Photon runtime can split long recordings at pauses. Ultra is tuned for accuracy, Redux compressed for size; both cover 25 European languages with automatic detection, punctuation, capitalisation and word timestamps. On Moondream’s own tests Ultra beats V3 on English, FLEURS, business speech (AMI, VoxPopuli, Earnings-22) and noisy audio. For Mac and iPhone dictation apps this is now the default pick; for a CPU-only server, Redux.

3. Canary-Qwen 2.5B (NVIDIA). Still the open-model leader on the English Open ASR Leaderboard at roughly 5.1–5.6% average WER depending on snapshot. It is English-only and batch-only, so it is the model you reach for when transcribing recorded English where every word matters — depositions, interviews, podcasts — and not for live use.

4. Whisper large-v3-turbo (OpenAI). Four decoder layers instead of 32 makes it about 8× faster than large-v3 at nearly the same accuracy, and whisper.cpp runs it on an 8 GB laptop or an iPhone. It has no native streaming and processes audio in 30-second windows, so long-form needs WhisperX or whisper.cpp’s VAD. It stays at number four because for anything outside the 13 Voxtral or 25 Parakeet languages it is the only serious option, and because three years of tooling (faster-whisper, WhisperX, MLX ports) is worth a lot in production.

5. Parakeet Unified 0.6B (NVIDIA). April 2026, English-only RNN-T model that does both whole-recording and streaming from one checkpoint. NVIDIA’s card reports 5.91% WER whole-recording and 6.29% streaming at 1.12 s latency, with latency configurable down to about 160 ms. The companion Parakeet Realtime EOU 120M streams at 80–160 ms and emits end-of-utterance markers for voice agents, at 9.30% WER and no punctuation.

6. Moonshine Tiny v2 (Useful Sensors). Built for streaming on constrained hardware; latency scales with audio length instead of padding to 30 seconds like Whisper, which is why it beats Whisper on time-to-first-token on short clips. English only. The right choice for a Raspberry Pi or a browser extension that needs live captions.

7. Whistle (Cactus Compute). The newest entry at 16.9 MB, a single file with no dependencies, seven languages, word timestamps with probabilities, keyword biasing, and 11 ms to first token on an M4 Pro. It beats the 145 MB Whisper Base on LibriSpeech, SPGISpeech, Earnings-22 and FLEURS while losing on TED-LIUM, AMI and MLS. Its unique trick is sharing a C++ engine with Cactus’s Needle tool-calling model so one binary turns a clip into function calls. Last on the list only because 30-second passes and seven languages limit the use cases; first on the list if your device has 20 MB to spare. Full comparison: Whistle vs Whisper Base vs Moonshine.

How to choose in 30 seconds

Need a language outside Voxtral’s 13 or Parakeet’s 25: Whisper large-v3-turbo. Nothing else covers it.
Live captions, multilingual, GPU available: Voxtral Realtime.
Dictation app on Mac/iPhone: Parakeet Ultra (or Redux for the smallest download).
Batch English transcription, accuracy above all: Canary-Qwen 2.5B.
Voice agent needing end-of-turn detection: Parakeet Realtime EOU 120M, then a bigger model for the final transcript.
Edge device under 50 MB: Whistle (multilingual, tool calls) or Moonshine Tiny (English streaming).

Read every WER with its benchmark attached. Vendor numbers on this page come from different test sets, normalisers and snapshots; a 0.5-point gap between two vendors’ self-reported averages is noise. The Open ASR Leaderboard is the one place where the same harness scores every model.

Last verified: October 4, 2026. WER figures are the vendors’ published numbers on the datasets named; the Open ASR Leaderboard is the cross-vendor reference.

The short answer

The ranking

Why each one is ranked where it is

How to choose in 30 seconds

Sources