QuartzNet models are end-to-end neural automatic speech recognition (ASR) models that transcribe segments of audio to text.
These models are based on the QuartzNet architecture, which is a variant of Jasper that uses 1D time-channel separable convolutional layers in its convolutional residual blocks and are therefore smaller than Jasper models.
QuartzNet models take in audio segments and transcribe them to letter, byte pair, or word piece sequences. The pretrained models here can be used immediately for fine-tuning or dataset evaluation.
Primary use case intended for these models is automatic speech recognition.
Input: Single-channel audio files (WAV) with a 16kHz sample rate
Output: Transcripts, which are sequences of valid vocabulary labels as given by the specification file
QuartzNet is an end-to-end architecture that is trained using CTC loss. These model checkpoints are intended to be used with the Train Adapt Optimize (TAO) Toolkit. In order to use these checkpoints, there should be a specification file (.yaml) that specifies hyperparameters, datasets for training and evaluation, and any other information needed for the experiment. For more information on the experiment spec files for each use case, please refer to the TAO Toolkit User Guide.
The model is encrypted and will only operate with the model encryption key
To fine-tune from a model checkpoint (.tlt), use the following command:
!tao speech-to-text finetune -e \ -m \ -g
Where the `` parameter should be a valid path to the file that specifies the fine-tuning hyperparameters, the dataset to fine-tune on, the dataset to evaluate on, and whether or not a change of vocabulary from the default (lowercase English letters, space, and apostrophe) is needed.
To evaluate an existing dataset using a model checkpoint (.tlt), use the following command:
!tao speech-to-text evaluate -e \ -m \ -g
The `` parameter should be a valid path to the file that specifies the dataset that is being evaluated.
This QuartzNet model was trained on a combination of seven datasets of English speech, with a total of 7,133 hours of audio samples. Samples were limited to a minimum duration of 0.1s long, and a maximum duration of 16.7s long. The model was trained for 300 epochs with Apex/Amp optimization level O1.
It achieves a Word Error Rate (WER) of 4.38% on LibriSpeech dev-clean, and a WER of 11.30% on LibriSpeech dev-other.
The datasets included in training are detailed in the table below.
|Wall Street Journal||81|
|Fisher English Training Speech||1,906|
|Mozilla Common Voice*||1,039|
|Appen English Speech||972|
|NSC Singapore English (Part 1)||1,857|
* Only non-dev and non-test validated clips from Mozilla Common Voice version en_1488h_2019-12-10.
Currently, TAO QuartzNet models only support training and inference on .wav audio files. All models included here were trained and evaluated on audio files with a sample rate of 16kHz, so for best performance you may need to upsample or downsample audio files to 16kHz.
In addition, the model will perform best on audio samples that are longer than 0.1 seconds long. For training and fine-tuning QuartzNet models, it is recommended that samples are capped at a maximum length of around 15 seconds, depending on the amount of memory available to you. You do not need to place a maximum length limitation for evaluation.
By downloading and using the models and resources packaged with TAO Conversational AI, you would be accepting the terms of the Riva license
NVIDIA’s platforms and application frameworks enable developers to build a wide array of AI applications. Consider potential algorithmic bias when choosing or creating the models being deployed. Work with the model’s developer to ensure that it meets the requirements for the relevant industry and use case; that the necessary instruction and documentation are provided to understand error rates, confidence intervals, and results; and that the model is being used under the conditions and in the manner intended.