nvidia/parakeet-tdt-0.6b-v2

Parakeet TDT 0.6B V2 (En)

| |

Description:

parakeet-tdt-0.6b-v2 is a 600-million-parameter automatic speech recognition (ASR) model designed for high-quality English transcription, featuring support for punctuation, capitalization, and accurate timestamp prediction. Try Demo here: https://huggingface.co/spaces/nvidia/parakeet-tdt-0.6b-v2

This XL variant of the FastConformer [1] architecture integrates the TDT [2] decoder and is trained with full attention, enabling efficient transcription of audio segments up to 24 minutes in a single pass. The model achieves an RTFx of 3380 on the HF-Open-ASR leaderboard with a batch size of 128. Note: RTFx Performance may vary depending on dataset audio duration and batch size.

Key Features

• Accurate word-level timestamp predictions
• Automatic punctuation and capitalization
• Robust performance on spoken numbers, and song lyrics transcription

For more information, refer to the Model Architecture section and the NeMo documentation.

This model is ready for commercial/non-commercial use.

License/Terms of Use:

GOVERNING TERMS: Use of this model is governed by the CC-BY-4.0 license.

Deployment Geography:

Global

Use Case:

This model serves developers, researchers, academics, and industries building applications that require speech-to-text capabilities, including but not limited to: conversational AI, voice assistants, transcription services, subtitle generation, and voice analytics platforms.

Release Date:

05/01/2025

Model Architecture:

Architecture Type:

FastConformer-TDT

Network Architecture:

• This model was developed based on FastConformer encoder architecture[1] and TDT decoder[2]
• This model has 600 million model parameters.

Input:

• Input Type(s): 16kHz Audio
• Input Format(s): .wav and .flac audio formats
• Input Parameters: 1D (audio signal)
• Other Properties Related to Input: Monochannel audio

Output:

• Output Type(s): Text
• Output Format: String
• Output Parameters: 1D (text)
• Other Properties Related to Output: Punctuations and Capitalizations included.

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

How to Use this Model:

To train, fine-tune or play with the model you will need to install NVIDIA NeMo. We recommend you install it after you've installed latest PyTorch version.

pip install -U nemo_toolkit['asr']

The model is available for use in the NeMo toolkit [3], and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.

Automatically instantiate the model

import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.ASRModel.from_pretrained(model_name="nvidia/parakeet-tdt-0.6b-v2")

Transcribing using Python

First, let's get a sample

wget https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav

Then simply do:

output = asr_model.transcribe(['2086-149220-0033.wav'])
print(output[0].text)

Transcribing with timestamps

To transcribe with timestamps:

output = asr_model.transcribe(['2086-149220-0033.wav'], timestamps=True)
# by default, timestamps are enabled for char, word and segment level
word_timestamps = output[0].timestamp['word'] # word level timestamps for first sample
segment_timestamps = output[0].timestamp['segment'] # segment level timestamps
char_timestamps = output[0].timestamp['char'] # char level timestamps

for stamp in segment_timestamps:
    print(f"{stamp['start']}s - {stamp['end']}s : {stamp['segment']}")

Software Integration:

Runtime Engine(s):

• NeMo 2.2

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
• NVIDIA Blackwell
• NVIDIA Hopper
• NVIDIA Volta

[Preferred/Supported] Operating System(s):

• Linux

Hardware Specific Requirements:

Atleast 2GB RAM for model to load. The bigger the RAM, the larger audio input it supports.

Model Version

Current version: parakeet-tdt-0.6b-v2. Previous versions can be accessed here.

Training and Evaluation Datasets:

Training

This model was trained using the NeMo toolkit [3], following the strategies below:

• Initialized from a wav2vec SSL checkpoint pretrained on the LibriLight dataset[7].
• Trained for 150,000 steps on 128 A100 GPUs.
• Dataset corpora were balanced using a temperature sampling value of 0.5.
• Stage 2 fine-tuning was performed for 2,500 steps on 4 A100 GPUs using approximately 500 hours of high-quality, human-transcribed data of NeMo ASR Set 3.0.

Training was conducted using this example script and TDT configuration.

The tokenizer was constructed from the training set transcripts using this script.

Training Dataset

The model was trained on the Granary dataset, consisting of approximately 120,000 hours of English speech data:

• 10,000 hours from human-transcribed NeMo ASR Set 3.0, including:

  • LibriSpeech (960 hours)
  • Fisher Corpus
  • National Speech Corpus Part 1
  • VCTK
  • VoxPopuli (English)
  • Europarl-ASR (English)
  • Multilingual LibriSpeech (MLS English) – 2,000-hour subset
  • Mozilla Common Voice (v7.0)
  • AMI

• 110,000 hours of pseudo-labeled data from:

  • YTC (YouTube-Commons) dataset[4]
  • YODAS dataset [5]
  • Librilight [7]

All transcriptions preserve punctuation and capitalization. The Granary dataset will be made publicly available after presentation at Interspeech 2025.

Data Collection Method by dataset

• Hybrid: Automated, Human

Labeling Method by dataset

• Hybrid: Synthetic, Human

Properties:

• Noise robust data from various sources
• Single channel, 16kHz sampled data

Evaluation Dataset

Huggingface Open ASR Leaderboard datasets are used to evaluate the performance of this model.

Data Collection Method by dataset

• Human

Labeling Method by dataset

• Human

Properties:

• All are commonly used for benchmarking English ASR systems.
• Audio data is typically processed into a 16kHz mono channel format for ASR evaluation, consistent with benchmarks like the Open ASR Leaderboard.

Performance

Huggingface Open-ASR-Leaderboard Performance

The performance of Automatic Speech Recognition (ASR) models is measured using Word Error Rate (WER). Given that this model is trained on a large and diverse dataset spanning multiple domains, it is generally more robust and accurate across various types of audio.

Base Performance

The table below summarizes the WER (%) using a Transducer decoder with greedy decoding (without an external language model):

Model	Avg WER	AMI	Earnings-22	GigaSpeech	LS test-clean	LS test-other	SPGI Speech	TEDLIUM-v3	VoxPopuli
parakeet-tdt-0.6b-v2	6.05	11.16	11.15	9.74	1.69	3.19	2.17	3.38	5.95

Noise Robustness

Performance across different Signal-to-Noise Ratios (SNR) using MUSAN music and noise samples:

SNR Level	Avg WER	AMI	Earnings	GigaSpeech	LS test-clean	LS test-other	SPGI	Tedlium	VoxPopuli	Relative Change
Clean	6.05	11.16	11.15	9.74	1.69	3.19	2.17	3.38	5.95	-
SNR 50	6.04	11.11	11.12	9.74	1.70	3.18	2.18	3.34	5.98	+0.25%
SNR 25	6.50	12.76	11.50	9.98	1.78	3.63	2.54	3.46	6.34	-7.04%
SNR 5	8.39	19.33	13.83	11.28	2.36	5.50	3.91	3.91	6.96	-38.11%

Telephony Audio Performance

Performance comparison between standard 16kHz audio and telephony-style audio (using μ-law encoding with 16kHz→8kHz→16kHz conversion):

Audio Format	Avg WER	AMI	Earnings	GigaSpeech	LS test-clean	LS test-other	SPGI	Tedlium	VoxPopuli	Relative Change
Standard 16kHz	6.05	11.16	11.15	9.74	1.69	3.19	2.17	3.38	5.95	-
μ-law 8kHz	6.32	11.98	11.16	10.02	1.78	3.52	2.20	3.38	6.52	-4.10%

These WER scores were obtained using greedy decoding without an external language model. Additional evaluation details are available on the Hugging Face ASR Leaderboard.[6]

References

[1] Fast Conformer with Linearly Scalable Attention for Efficient Speech Recognition

[2] Efficient Sequence Transduction by Jointly Predicting Tokens and Durations

[3] NVIDIA NeMo Toolkit

[4] Youtube-commons: A massive open corpus for conversational and multimodal data

[5] Yodas: Youtube-oriented dataset for audio and speech

[6] HuggingFace ASR Leaderboard

[7] MOSEL: 950,000 Hours of Speech Data for Open-Source Speech Foundation Model Training on EU Languages

Inference:

Engine:

• NVIDIA NeMo

Test Hardware:

• NVIDIA A10
• NVIDIA A100
• NVIDIA A30
• NVIDIA H100
• NVIDIA L4
• NVIDIA L40
• NVIDIA Turing T4
• NVIDIA Volta V100

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their supporting model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards here.

Please report security vulnerabilities or NVIDIA AI Concerns here.

Bias:

Field	Response
Participation considerations from adversely impacted groups protected classes in model design and testing	None
Measures taken to mitigate against unwanted bias	None

Explainability:

Field	Response
Intended Domain	Speech to Text Transcription
Model Type	FastConformer
Intended Users	This model is intended for developers, researchers, academics, and industries building conversational based applications.
Output	Text
Describe how the model works	Speech input is encoded into embeddings and passed into conformer-based model and output a text response.
Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of	Not Applicable
Technical Limitations & Mitigation	Transcripts may be not 100% accurate. Accuracy varies based on language and characteristics of input audio (Domain, Use Case, Accent, Noise, Speech Type, Context of speech, etc.)
Verified to have met prescribed NVIDIA quality standards	Yes
Performance Metrics	Word Error Rate
Potential Known Risks	If a word is not trained in the language model and not presented in vocabulary, the word is not likely to be recognized. Not recommended for word-for-word/incomplete sentences as accuracy varies based on the context of input text
Licensing	GOVERNING TERMS: Use of this model is governed by the CC-BY-4.0 license.

Privacy:

Field	Response
Generatable or reverse engineerable personal data?	None
Personal data used to create this model?	None
Is there provenance for all datasets used in training?	Yes
Does data labeling (annotation, metadata) comply with privacy laws?	Yes
Is data compliant with data subject requests for data correction or removal, if such a request was made?	No, not possible with externally-sourced data.
Applicable Privacy Policy	https://www.nvidia.com/en-us/about-nvidia/privacy-policy/

Safety:

Field	Response
Model Application(s)	Speech to Text Transcription
Describe the life critical impact	None
Use Case Restrictions	Abide by CC-BY-4.0 License
Model and dataset restrictions	The Principle of least privilege (PoLP) is applied limiting access for dataset generation and model development. Restrictions enforce dataset access during training, and dataset license constraints adhered to.