whatmodel.sh / stt / Parakeet TDT 0.6B v3

Parakeet TDT 0.6B v3

+ contribute

Parakeet TDT 0.6B v3 is an open-weights speech-to-text model from NVIDIA, released August 2025. Its published size is 0.63B parameters. Licence: cc-by-4.0.

56.9±1.1#56 of 162 speech-to-text modelsfrom 39 benchmark results across 17 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 v3, out of 176 with a score, then more from NVIDIA.

#	model	score	where	RAM	released
63	Fun-ASR-MLT-Nano-2512 Alibaba (FunAudioLLM)	57±4	local	~0.9GB	2025-12
64	Parakeet Redux Moondream	57±4	local	~0.2GB	2026-09
65	Parakeet TDT 0.6B v3 NVIDIA	57±1	api $0.0015/audio min +local	~2GB	2025-08
66	Kyutai STT 2.6B (en) Kyutai	57±4	local	~6.7GB	2025-06
67	Whisper large-v3 OpenAI	57±1	api $0.0018/audio min +local	~4.2GB	2023-11
	more from NVIDIA
28	Canary-Qwen-2.5B NVIDIA	61±2	local	~6.6GB	2025-07
35	Parakeet TDT 0.6B v2 NVIDIA	60±1	local	~1.9GB	2025-05
37	Parakeet Unified EN 0.6B NVIDIA	60±4	local	~0.7GB	2026-04

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

$ what it's good at

long-form transcription▁▁▄▁·▁▂█▄▁▁▁ #9 of 27 * · 3.8% of its score · 1 test

Beats Chirp 3, Parakeet TDT 1.1B, Parakeet RNNT 1.1B here, though they rank higher overall. Trails Rev AI Machine (Reverb ASR), Whisper large-v3 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Open ASR English long-form WER	9.1% ✎	▁▁▅▁·▃▅█▆▁▂▂ beats 19/30	avg 12%	▇▇ 3.8%	typical

multilingual transcription▂··▃▄█▆▅▄▂·▂ #17 of 36 · 17% of its score · 3 tests

Beats Qwen3-ASR-1.7B, Voxtral Mini Transcribe V2, Canary-1B-v2 here, though they rank higher overall. Trails Omnilingual ASR CTC 1B v2, SeamlessM4T v2 Large, Omnilingual ASR LLM 1B v2 here, though they rank lower overall.

test	result	vs the others in its run	difficulty	counts	why
Open ASR multilingual WER	4.8% ✎	▂··▂▃▂█▃█▂▁▁ beats 19/29	avg 8.9%	▇▇▇▇▇ 13%	typical
CoSE-E enterprise code-switching · 4 European pairs mean WER	8.6% ✎	█··██···█··█ beats 0/4	avg 13%	▇ 3.6%	typical
Iberian ASR benchmark · 7-language macro WER (eu, ca, gl, pt, es + de, tr)	51.8% ✎	█▃▃··▅▃▃··▃▃ beats 2/10	avg 38%	▇ 0.7%	noisy

other tests30 not in a task · 78% of its score

test	result	vs the others in its run	difficulty	counts	why
FLEURS 25 European languages, mean WER (NVIDIA set) +2	11.6% ✎	█····█·····█ beats 0/2	avg 18%	▇▇▇ 6.3%	small run · several runs
LibriSpeech test-clean WER +1	1.9% ✎	▁···▁▁▂·▄▄██ beats 27/56	avg 2.3%	▇▇ 5.6%	blunt · several runs
Mega-ASR adverse-condition suite · mean WER (CHiME-4, VOiCES, NOIZEUS)	11.2% ✎	▄····█▄▄···▄ beats 3/5	avg 19%	▇▇ 4.3%	typical
SPGISpeech test WER lab	4.0% ✎	▄▄··▄██··▄██ beats 1/11	avg 5.4%	▇▇ 3.9%	sharp
Open ASR English short-form WER (March 2026)	6.3% ✎	▁▁▁▃··▁▁▂██▆ beats 74/86	avg 8.2%	▇▇ 3.9%	blunt
CoVoST 2 source speech, 13 languages, mean ASR WER (NVIDIA set) lab +1	12.0% ✎	█·····█····█ beats 0/2	avg 17%	▇▇ 3.8%	small run · several runs
Multilingual LibriSpeech 6 languages, mean WER (NVIDIA set) lab +1	7.8% ✎	█·▄········▄ beats 1/3	avg 11%	▇ 3.6%	several runs
VoiceCodeBench · batch canonical token/entity match (CTEM)	78.6% ✎	▄···▄··▄▄▄·█ beats 0/6	easy · avg 77%	▇ 3.5%	typical

22 more tests

test	result	vs the others in its run	difficulty	counts	why
Earnings-22 WER (short-form, Open ASR / ESB test split) lab	11.4% ✎	█······███·█ beats 2/4	avg 21%	▇ 3.3%	typical
AMI WER (Open ASR / ESB test split) lab	11.3% ✎	▄▄·······█·▄ beats 2/4	avg 24%	▇ 3.1%	typical
RW-Voice-EQ ASR · 4 real-world English sets, mean WER	8.3% ✎	█▃▃▅▃···▃··▃ beats 7/9	avg 15%	▇ 3.1%	typical
GigaSpeech test WER lab	9.6% ✎	▁····▁··▁▄█▂ beats 14/17	avg 17%	▇ 2.9%	typical
Dutch ASR Leaderboard · average WER (FLEURS, VoxPopuli, MLS nl)	10.1% ✎	▄··▄█···█··▄ beats 3/6	avg 16%	▇ 2.8%	typical
TED-LIUM long-form WER	2.7% ✎	█····█·█████ beats 4/6	avg 4.6%	▇ 2.7%	sharp · small run
LibriSpeech test-other WER lab	3.6% ✎	▁▁▁▂▅▃▅▄█▃▂▁ beats 47/54	avg 5.4%	▇ 2.3%	typical
Granite 5.0 report · Open ASR public 7-set mean WER	6.3% ✎	█···█······█ beats 1/2	avg 11%	▇ 2.3%	small run
Russian ASR Leaderboard (N. Shmyrev test sets) · 11-dataset mean WER	14.8% ✎	█·····█····█ beats 2/2	avg 25%	▇ 2.3%	small run
Russian STT · podlodka corpus WER (20-clip pilot)	9.3% ✎	█··········▄ beats 0/2	avg 16%	▇ 2.2%	small run
Russian STT · rulibrispeech corpus WER	10.8% ✎	█··█·······█ beats 0/2	avg 19%	▇ 2.2%	small run
VoiceCodeBench · batch WER (300 English workplace segments)	9.3% ✎	▄·▄·▄·▄█···▄ beats 0/6	avg 12%	▇ 2.1%	typical
FLEURS English test WER +1	4.8% ✎	▁··▁▁▁▁▁▁▃█▁ beats 28/56	avg 5.8%	▇ 2.0%	blunt
Open ASR Leaderboard – Chilean Spanish · 3-dataset average WER	7.9% ✎	▂▂·····▂▂·█▂ beats 7/8	avg 13%	▇ 1.8%	typical
Russian STT · fleurs corpus WER	5.6% ✎	█··█·······█ beats 1/2	avg 9.9%	▇ 1.7%	small run
VoxPopuli English test WER lab	6.1% ✎	█▄▄··████▄██ beats 11/16	avg 9.0%	▇ 1.6%	typical
Danish ASR Leaderboard · mean WER over 5 Danish test sets (RyeAI)	30.7% ✎	▂▂···▃▂▅█▃▂▂ beats 6/16	avg 32%	▇ 1.5%	blunt
Betthupferl (German dialect ASR benchmark) · Standard German audio → Standard German transcript WER	15.6% ✎	▄·█·▄▄▄▄·█▄▄ beats 6/10	avg 15%	▇ 1.1%	blunt
verli STT benchmark · SPGISpeech 1,200-clip sample WER	3.8% ✎	▂·····▄▂█▂▂▄ beats 2/11	avg 4.6%	▇ 1.1%	typical
verli STT benchmark · Earnings-22 WER (2,741 utt.)	11.2% ✎	▃·▃▅··▃▃▅▅█▃ beats 9/13	avg 16%	▇ 1.0%	blunt
Betthupferl (German dialect ASR benchmark) · dialect audio → Standard German transcript WER	43.1% ✎	▃█····▃▅▃·▅▃ beats 6/10	avg 41%	▇ 0.5%	blunt
TED-LIUM test WER (short-form) lab	2.8% ✎	▁·▁▇█▃·▁▁▃█▄ beats 26/27	avg 5.0%	▇ 0.1%	blunt

$ where to run it

Through an API

api · Together AI · $0.0015/audio min · www.together.ai ↗ · price sourceUSD per input audio minute: Together lists NVIDIA Parakeet TDT 0.6B v3 at $0.0015 in its 'Price per audio minute' table (serverless transcription; no separate batch price shown). Product: Together serverless audio transcription. API model: nvidia/parakeet-tdt-0.6b-v3. Settings: standard. Price basis: published list rate. Exclusions: 'Realtime' variant. Currency: USD. Together official pricing page, observed 2026-09-25.

Other ways to run it

version

Run it as published · needs ~2.01 GB, Mac only · from github.com/senstella
uvx parakeet-mlx audio.mp3

Official weights and documentation: huggingface.co/nvidia ↗

$ how the score was worked out · NVIDIA ASR model cards (Parak… 22%, Open ASR Leaderboard 21%, Moondream parakeet-ultra and… 7% and 14 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	11	▇▇▇▇▇▇▇ 22%	CC BY 4.0
Open ASR Leaderboard	3	▇▇▇▇▇▇ 21%	Apache-2.0
Moondream parakeet-ultra and parakeet-redux model cards (Benchmarks sections)	3	▇▇ 7%	CC BY 4.0
transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports	2	▇▇ 7%	MIT
Canary-1B-v2 & Parakeet-TDT-0.6B-v3: Efficient and High-Performance Models for Multilingual ASR and AST (NVIDIA, arXiv:2509.14128v2) · lab-reported	3	▇▇ 7%	CC BY-SA 4.0
Russian STT benchmark by domain	3	▇▇ 6%	MIT
VoiceCodeBench (arXiv:2608.28916v2) · structured-token recovery ASR benchmark	2	▇▇ 6%	CC BY 4.0
Mega-ASR (arXiv:2605.19833v1) · adverse-condition ASR baselines (CHiME-4, VOiCES, NOIZEUS)	1	▇ 4%	CC BY 4.0
CoSE-E: A Benchmark for Code-switched Speech Evaluation in Enterprise Settings (arXiv:2609.35645v1), Table 2	1	▇ 4%	CC BY 4.0
RW-Voice-EQ Bench (Hume AI)	1	▇ 3%	CC BY 4.0
Dutch ASR Leaderboard (tvosch) · committed result files	1	▇ 3%	Apache-2.0
Russian ASR Leaderboard (ESpeech; benchmark by Nikolay Shmyrev)	1	▇ 2%	MIT
verliapp/stt-benchmark · independent STT accuracy benchmark	2	▇ 2%	MIT
Open ASR Leaderboard – Chilean Spanish (idsudd / Alonso Astroza)	1	▇ 2%	Apache-2.0
German dialect ASR Benchmark (Betthupferl; kamaal-08/bavarian-asr-leaderboard)	2	▇ 2%	MIT
Danish Speech-to-Text Leaderboard (RyeAI) · results dataset	1	▇ 1%	MIT
Benchmarking Automatic Speech Recognition Tools for Iberian Languages (arXiv:2609.36920v1), Table 3	1	▇ 1%	CC BY 4.0
prior · pull toward the middle		▇ 0%
weighted average		= 56.9

not counted (1)

NVIDIA ASR model cards (Parakeet / Canary, Hugging Face)

FLEURS English test WER [correct]

4.9%

the only model in this source's run of the test, so there is nothing to compare it with there; another run of the test counts

$ 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)	35.4 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	31.4 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)	91.4 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)	57.5 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	290.8 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	318.5 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	99.1 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	232.3 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 ↗

$ 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