Open source speech to text models

The six, side by side

Every cell comes from the model card, the repo README or the project page. Turbo is 809 M because it is large-v3 with the decoder cut from 32 layers to 4. Canary is 978 million despite the 1B in its name.

Fraze is not open source. It is in the table because it ships one of these models, Parakeet v3, picked for 25 European languages in 600 M parameters and for how it held up on English call audio in our runs below. On macOS 26 it can switch to Apple's SpeechAnalyzer for languages Parakeet lacks, such as Japanese, Korean and Chinese.

Which one fits which job

• Whisper large-v3 and large-v3-turbo: 99 languages listed. Large-v3 can also translate speech into English. The repo says turbo is not trained for translation. It lists about 10 GB of VRAM for large and about 6 GB for turbo.
• Parakeet TDT 0.6B v3: automatic language detection, word and segment timestamps. The card targets NVIDIA GPUs on Linux. NVIDIA also ships a quantised GGUF file and the NeMo-Speech.cpp runtime, which has CPU, Metal, Vulkan and CUDA builds and needs no Python.
• Canary 1B v2: the same 25 languages as Parakeet, plus speech translation from English into 24 and from 24 into English. Its card reports 7.15 average word error rate at 749 RTFx on the Open ASR Leaderboard, and targets NVIDIA GPUs on Linux.
• Moonshine: one library for Python, JavaScript, iOS, Android, desktop systems and Raspberry Pi. The docs report 6.65 percent average word error rate for English Medium Streaming, its largest model at 245 M. The old usefulsensors/moonshine address redirects to moonshine-ai/moonshine.
• Kyutai STT: the 1B model has a semantic voice activity detector, so far only in the Rust server. The project page says an H100 can carry 400 real time streams at once. There is an MLX build for Mac and iPhone.
• WhisperKit, from Argmax: not a model but a way to run Whisper on Apple hardware through Core ML. For multilingual accuracy Argmax recommends large-v3-v20240930_626MB, a compressed large-v3-turbo. The old argmaxinc/WhisperKit address redirects to argmaxinc/argmax-oss-swift.

How to run each one

These are the entry points the projects document themselves. Whisper also needs ffmpeg. The Parakeet commands need the NeMo-Speech.cpp runtime installed first. The Canary line installs NeMo, after which you load the model in Python. The Kyutai sample file is in its delayed-streams-modeling repo.

# Whisper, the reference implementation
pip install -U openai-whisper
whisper audio.flac audio.mp3 audio.wav --model turbo

# Parakeet TDT 0.6B v3, native C++ runtime, no Python
hf download nvidia/parakeet-tdt-0.6b-v3 \
  parakeet-tdt-0.6b-v3.q8_0.gguf \
  --local-dir models
nemo-speech transcribe audio.wav \
  --model models/parakeet-tdt-0.6b-v3.q8_0.gguf

# Canary 1B v2, through NVIDIA NeMo
pip install -U nemo_toolkit['asr']

# Moonshine, live from a microphone
pip install moonshine-voice
moonshine-voice mic --language en

# Kyutai STT on a file
uvx --with moshi python -m moshi.run_inference \
  --hf-repo kyutai/stt-2.6b-en audio/bria.mp3

# WhisperKit on a Mac
brew install whisperkit-cli

What the Open ASR Leaderboard shows

The Hugging Face Open ASR Leaderboard is the public comparison most of these cards point at. It ranks by average word error rate, reports RTFx, a speed measure where higher is better, and has separate tabs for multilingual and long-form results.

Two warnings. It scores proprietary API models next to open ones, so the top of the table is not a list of things you can download. And it changes: the page said it was last updated on 18 September 2026, and its default average covered ten datasets, two of them private. The Canary and Moonshine averages above cover eight, so they will not match the live table.

Our own runs, on the same clips

• LibriSpeech clean: Apple SpeechAnalyzer 1.82, Whisper large-v3 1.82, Whisper large-v3-turbo 1.93, Parakeet TDT 0.6B v3 2.15, a cloud engine 2.78. The top models sit inside one another's confidence intervals, and LibriSpeech is in Parakeet's training data.
• Earnings-22 conference calls, in no engine's listed training data: Parakeet v3 11.20, the cloud engine 11.61, Apple 12.03, Whisper large-v3-turbo run on-device through WhisperKit 12.35. The benchmark repository says the reference PyTorch turbo scores about 11.1 on the same set, level with Parakeet. Parakeet and the cloud engine overlap too, so treat those three as a tie.
• FLEURS read speech, 200 clips per language, measured on 2026-08-30. Whisper here is large-v3-turbo, hosted, not run on-device. German: Whisper 5.33, Parakeet 6.21, Apple 7.37. Spanish: 5.68, 6.05, 7.04. French: 6.82, 7.83, 8.42. Russian: Whisper 7.48, Parakeet 9.18, Apple not supported.
• Speed: Parakeet v3 ran at about 0.02 times real time on an Apple silicon Mac, so about 50 times faster than the audio. We did not rank the engines on speed.

Our own measurements, not model card figures. Word error rate in percent, lower is better. Whisper ran on-device through WhisperKit, not OpenAI's reference PyTorch build, unless marked hosted. Parakeet ran through parakeet-mlx, not the int8 build Fraze ships.

Reading those numbers honestly

On call audio Parakeet is level with reference Whisper and the cloud engine, and ahead of Whisper as it runs through WhisperKit. On clean speech and on all four non-English sets Whisper scored lower. With 200 clips per language the intervals are wide: hosted Whisper's German interval runs from 4.22 to 6.68, which takes in Parakeet's 6.21.

Whisper also lists 99 languages to Parakeet's 25, so for Hindi, Arabic or Thai start there. For a model that does not wait for the end of a sentence, start with Kyutai STT or Moonshine. Parakeet's case is narrower: 25 European languages in one 600 M model. That is the case Fraze needed. Yours may differ.

Questions