parakeet-tdt_ctc-0.6b-ja is an ASR model that transcribes Japanese speech with Punctuations. This model is developed by NVIDIA NeMo team.
It is an XL version of Hybrid FastConformer [1] TDT-CTC [2] (around 0.6B 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['asr']
How to Use this Model
The model is available for use in the NeMo Framework [3], and can be used as a pre-trained checkpoint for inference or for fine-tuning on another dataset.
Automatically instantiate the model
python
1import nemo.collections.asr as nemo_asr
2asr_model = nemo_asr.models.ASRModel.from_pretrained(model_name="nvidia/parakeet-tdt_ctc-0.6b-ja")
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 Hybrid FastConformer-TDT-CTC 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_ctc-0.6b-ja 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 Framework [3] was used for training this model with this example script and this base config.
The model was trained for 300k steps with dynamic bucketing and a batch duration of 600s per GPU on 32 NVIDIA A100 80GB GPUs, and then finetuned for 100k additional steps on the modified training data (predicted texts for training samples with CER>10%).
SentencePiece [4] tokenizer with 3072 tokens for this model was built using the text transcripts of the train set with this script.
Datasets
The model was trained on ReazonSpeech v2.0 [5] speech corpus containing more than 35k hours of natural Japanese speech.
Performance
The following table summarizes the performance of this model in terms of Character Error Rate (CER%).
In CER calculation, punctuation marks and non-alphabet characters are removed, and numbers are transformed to words using num2words library [6].
Version
Decoder
JSUT basic5000
MCV 8.0 test
MCV 16.1 dev
MCV16.1 test
TEDxJP-10k
1.23.0
TDT
6.4
7.1
10.1
13.2
9.0
1.23.0
CTC
6.5
7.2
10.2
13.3
9.1
These are greedy CER numbers without external LM.
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
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.
