Parakeet TDT 1.1B (en)

| |

parakeet-tdt-1.1b is an ASR model that transcribes speech in lower case English alphabet. This model is jointly developed by NVIDIA NeMo and Suno.ai teams. It is an XXL version of FastConformer [1] TDT [2] (around 1.1B parameters) model. See the model architecture section and NeMo documentation for complete architecture details.

NVIDIA NeMo: Training

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 nemo_toolkit['all']

How to Use this Model

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.EncDecRNNTBPEModel.from_pretrained(model_name="nvidia/parakeet-tdt-1.1b")

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:

asr_model.transcribe(['2086-149220-0033.wav'])

Transcribing many audio files

python [NEMO_GIT_FOLDER]/examples/asr/transcribe_speech.py 
 pretrained_name="nvidia/parakeet-tdt-1.1b" 
 audio_dir="<DIRECTORY CONTAINING AUDIO FILES>"

Input

This model accepts 16000 Hz mono-channel audio (wav files) as input.

Output

This model provides transcribed speech as a string for a given audio sample.

Model Architecture

This model uses a FastConformer-TDT architecture. FastConformer [1] is an optimized version of the Conformer model with 8x depthwise-separable convolutional downsampling. You may find more information on the details of FastConformer here: Fast-Conformer Model.

TDT (Token-and-Duration Transducer) [2] is a generalization of conventional Transducers by decoupling token and duration predictions. Unlike conventional Transducers which produces a lot of blanks during inference, a TDT model can skip majority of blank predictions by using the duration output (up to 4 frames for this parakeet-tdt-1.1b model), thus brings significant inference speed-up. The detail of TDT can be found here: Efficient Sequence Transduction by Jointly Predicting Tokens and Durations.

Training

The NeMo toolkit [3] was used for training the models for over several hundred epochs. These model are trained with this example script and this base config.

The tokenizers for these models were built using the text transcripts of the train set with this script.

Datasets

The model was trained on 64K hours of English speech collected and prepared by NVIDIA NeMo and Suno teams.

The training dataset consists of private subset with 40K hours of English speech plus 24K hours from the following public datasets:

Librispeech 960 hours of English speech
Fisher Corpus
Switchboard-1 Dataset
WSJ-0 and WSJ-1
National Speech Corpus (Part 1, Part 6)
VCTK
VoxPopuli (EN)
Europarl-ASR (EN)
Multilingual Librispeech (MLS EN) - 2,000 hour subset
Mozilla Common Voice (v7.0)
People's Speech - 12,000 hour subset

Performance

The performance of Automatic Speech Recognition models is measuring using Word Error Rate. Since this dataset is trained on multiple domains and a much larger corpus, it will generally perform better at transcribing audio in general.

The following tables summarizes the performance of the available models in this collection with the Transducer decoder. Performances of the ASR models are reported in terms of Word Error Rate (WER%) with greedy decoding.

Version	Tokenizer	Vocabulary Size	AMI	Earnings-22	Giga Speech	LS test-clean	SPGI Speech	TEDLIUM-v3	Vox Populi	Common Voice
1.22.0	SentencePiece Unigram	1024	15.90	14.65	9.55	1.39	2.62	3.42	3.56	5.48

These are greedy WER numbers without external LM. More details on evaluation can be found at HuggingFace ASR Leaderboard

Model Fairness Evaluation

As outlined in the paper "Towards Measuring Fairness in AI: the Casual Conversations Dataset", we assessed the parakeet-tdt-1.1b model for fairness. The model was evaluated on the CausalConversations-v1 dataset, and the results are reported as follows:

Gender Bias:

Gender	Male	Female	N/A	Other
Num utterances	19325	24532	926	33
% WER	17.18	14.61	19.06	37.57

Age Bias:

Age Group	$(18-30)$	$(31-45)$	$(46-85)$	$(1-100)$
Num utterances	15956	14585	13349	43890
% WER	15.83	15.89	15.46	15.74

(Error rates for fairness evaluation are determined by normalizing both the reference and predicted text, similar to the methods used in the evaluations found at https://github.com/huggingface/open_asr_leaderboard.)

NVIDIA Riva: Deployment

NVIDIA Riva, is an accelerated speech AI SDK deployable on-prem, in all clouds, multi-cloud, hybrid, on edge, and embedded. Additionally, Riva provides:

World-class out-of-the-box accuracy for the most common languages with model checkpoints trained on proprietary data with hundreds of thousands of GPU-compute hours
Best in class accuracy with run-time word boosting (e.g., brand and product names) and customization of acoustic model, language model, and inverse text normalization
Streaming speech recognition, Kubernetes compatible scaling, and enterprise-grade support

Although this model isn’t supported yet by Riva, the list of supported models is here.
Check out Riva live demo.

References

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

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

[3] Google Sentencepiece Tokenizer

[4] NVIDIA NeMo Toolkit

[5] Suno.ai

[6] HuggingFace ASR Leaderboard

[7] Towards Measuring Fairness in AI: the Casual Conversations Dataset

Licence

License to use this model is covered by the CC-BY-4.0. By downloading the public and release version of the model, you accept the terms and conditions of the CC-BY-4.0 license.

nvidia
/

parakeet-tdt-1.1b