whatmodel.sh / stt / Parakeet TDT 0.6B v2

Parakeet TDT 0.6B v2

+ contribute

Parakeet TDT 0.6B v2 is an open-weights speech-to-text model from NVIDIA, released May 2025. Its published size is 0.6B parameters. Licence: cc-by-4.0.

60.4±1.4#30 of 162 speech-to-text modelsfrom 21 benchmark results across 7 sources

Score within the Open ASR English short-form WER (March 2026) comparison set only.

$ how it compares

The speech-to-text models ranked either side of Parakeet TDT 0.6B v2, out of 176 with a score, then more from NVIDIA.

#	model	score	where	RAM	released
33	Gemini 3.5 Transcribe Live Google	61±3	api $0.0090/audio min		2026-08
34	Granite Speech 4.1 2B Plus IBM	61±4	local	~2.4GB	2026-04
35	Parakeet TDT 0.6B v2 NVIDIA	60±1	local	~1.9GB	2025-05
36	Tencent Hy-ASR-3.0-preview Tencent	60±4	comparison only
37	Parakeet Unified EN 0.6B NVIDIA	60±4	local	~0.7GB	2026-04
	more from NVIDIA
28	Canary-Qwen-2.5B NVIDIA	61±2	local	~6.6GB	2025-07
40	Canary-1B-Flash NVIDIA	60±2	local	~4.4GB	2025-03
49	Canary-1B NVIDIA	59±3	local	~5.3GB	2024-02

see all speech-to-text models · find the ones that fit your machine

$ what it's good at

long-form transcription▁▁▄▁·▁▂█▄▁▁▁ #8 of 27 * · 6.7% of its score · 1 test

Beats Chirp 3 here, though they rank higher overall. Trails Rev AI Machine (Reverb ASR), Whisper large-v3, Whisper large-v3-turbo here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Open ASR English long-form WER	9.0% ✎	▁▁▅▁·▃▅█▆▁▂▂ beats 20/30	avg 12%	▇▇▇ 6.7%	typical

other tests15 not in a task · 93% of its score

test	result	vs the others in its run	difficulty	counts	why
AfriSpeech-MultiBench · 7-region African-accent mean WER	22.5% ✎	▂····▂···▂▆█ beats 8/11	avg 46%	▇▇▇▇▇ 13%	typical
LibriSpeech test-clean WER +2	1.7% ✎	▁···▁▁▂·▄▄██ beats 39/56	avg 2.3%	▇▇▇▇▇ 11%	blunt · several runs
AfriSpeech-MultiBench · Parliament (legal domain) WER	18.4% ✎	▂▂····▂·▅·▃█ beats 10/12	avg 52%	▇▇▇▇ 11%	sharp
LibriSpeech test-other WER +1	3.1% ✎	███·····██·█ beats 5/5	avg 5.7%	▇▇▇▇ 9.2%	several runs
SPGISpeech test WER lab	2.2% ✎	▄▄··▄██··▄██ beats 9/11	avg 5.4%	▇▇▇ 6.9%	sharp
Open ASR English short-form WER (March 2026)	6.0% ✎	▁▁▁▃··▁▁▂██▆ beats 77/86	avg 8.2%	▇▇▇ 6.8%	blunt
Earnings-22 WER (short-form, Open ASR / ESB test split) lab	11.2% ✎	█······███·█ beats 3/4	avg 21%	▇▇ 5.7%	typical
AMI WER (Open ASR / ESB test split) lab	11.2% ✎	▄▄·······█·▄ beats 3/4	avg 24%	▇▇ 5.4%	typical

7 more tests

test	result	vs the others in its run	difficulty	counts	why
VoxPopuli English test WER +1	5.3% ✎	▄··▄··▄···█▄ beats 3/5	avg 9.0%	▇▇ 5.1%	several runs
GigaSpeech test WER lab	9.7% ✎	▁····▁··▁▄█▂ beats 13/17	avg 17%	▇▇ 5.0%	typical
AfriCodeSwitch English-Yoruba code-switched speech · WER (cap2000)	66.1% ✎	█··█·█···█·█ beats 4/4	avg 131%	▇▇ 5.0%	typical
FLEURS English test WER +1	4.1% ✎	▁··▁▁▁▁▁▁▃█▁ beats 50/56	avg 5.8%	▇▇ 4.8%	blunt · several runs
verli STT benchmark · SPGISpeech 1,200-clip sample WER	2.0% ✎	▂·····▄▂█▂▂▄ beats 10/11	avg 4.6%	▇ 1.9%	typical
verli STT benchmark · Earnings-22 WER (2,741 utt.)	11.2% ✎	▃·▃▅··▃▃▅▅█▃ beats 8/13	avg 16%	▇ 1.7%	blunt
TED-LIUM test WER (short-form) lab	3.4% ✎	▁·▁▇█▃·▁▁▃█▄ beats 25/27	avg 5.0%	▇ 0.2%	blunt

$ where to run it

Other ways to run it

version

Run it as published · needs ~1.94 GB, Mac only · from github.com/senstella
uvx parakeet-mlx audio.mp3 --model mlx-community/parakeet-tdt-0.6b-v2

Official weights and documentation: huggingface.co/nvidia ↗

$ how the score was worked out · NVIDIA ASR model cards (Parak… 31%, AfriSpeech-MultiBench (arXiv:… 24%, Open ASR Leaderboard 13% and 4 more

Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works

source	results	share of the score	licence
NVIDIA ASR model cards (Parakeet / Canary, Hugging Face) · lab-reported	8	▇▇▇▇▇▇▇▇▇ 31%	CC BY 4.0
AfriSpeech-MultiBench (arXiv:2511.14255v1) · African-accented English ASR benchmark suite	2	▇▇▇▇▇▇▇ 24%	CC BY 4.0
Open ASR Leaderboard	2	▇▇▇▇ 13%	Apache-2.0
transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports	2	▇▇▇▇ 12%	MIT
microsoft/VibeVoice-ASR-BitNet model card (Accuracy table)	4	▇▇▇ 10%	MIT
Switch-aware evaluation of ASR on English-Yoruba code-switched speech (arXiv:2609.11786v1)	1	▇▇ 5%	CC BY 4.0
verliapp/stt-benchmark · independent STT accuracy benchmark	2	▇ 4%	MIT
prior · pull toward the middle		▇ 1%
weighted average		= 60.4

$ published speed measurements · 8

Reference results for the hardware, engine and settings below. They aren't verified speeds of the downloads above or predictions for your machine.

AMD Ryzen 7 9800X3D	onnx-default (precision not stated in table)	36.8 x-realtime	ONNX Runtime CPUExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16.	source ↗
AMD Ryzen 7 9800X3D	onnx-int8	30.5 x-realtime	ONNX Runtime CPUExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16.	source ↗
NVIDIA GeForce RTX 5070 Ti	onnx-default (precision not stated in table)	88.7 x-realtime	ONNX Runtime CUDAExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16.	source ↗
NVIDIA T4	onnx-default (precision not stated in table)	57.6 x-realtime	ONNX Runtime CUDAExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16.	source ↗
NVIDIA GeForce RTX 5070 Ti	TensorRT default precision	298.3 x-realtime	ONNX Runtime TensorrtExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16.	source ↗
NVIDIA GeForce RTX 5070 Ti	TensorRT fp16 enabled	329.2 x-realtime	ONNX Runtime TensorrtExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16.	source ↗
NVIDIA T4	TensorRT default precision	96 x-realtime	ONNX Runtime TensorrtExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16.	source ↗
NVIDIA T4	TensorRT fp16 enabled	237.2 x-realtime	ONNX Runtime TensorrtExecutionProvider · English suite; VoxPopuli en test; 100 samples used for warmup; per-waveform recognition timing; RTFx = summed audio duration / summed recognition time Exact published English table cell. Default precision is not silently relabeled fp32. TensorRT fp16 enables that provider setting; it does not assert every operation uses fp16.	source ↗

$ improve this page

Update model information Add a place to run it Add a test result Propose a benchmark Correct a benchmark score Edit data on GitHub

Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.

data updated 2026-10-11 02:20 UTC · sources and licences · json