automatic-speech-recognition speech audio CTC Conformer Transformer pytorch NeMo hf-asr-leaderboard Riva

NVIDIA Conformer-CTC Large (Kinyarwanda)

<style> img { display: inline; } </style>

| Model architecture | Model size | Language | Riva Compatible |

This model transcribes speech into lowercase Latin alphabet including spaces, and apostroph, and is trained on around 2000 hours of Kinyarwanda speech data. It is a non-autoregressive "large" variant of Conformer, with around 120 million parameters. See the model architecture section and NeMo documentation for complete architecture details. It is also compatible with NVIDIA Riva for production-grade server deployments.

Usage

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.

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']

Automatically instantiate the model

import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.EncDecCTCModelBPE.from_pretrained("nvidia/stt_rw_conformer_ctc_large")

Transcribing using Python

Simply do:

asr_model.transcribe(['<your_audio>.wav'])

Transcribing many audio files

python [NEMO_GIT_FOLDER]/examples/asr/transcribe_speech.py 
 pretrained_name="nvidia/stt_rw_conformer_ctc_large" 
 audio_dir="<DIRECTORY CONTAINING AUDIO FILES>"

Input

This model accepts 16 kHz mono-channel Audio (wav files) as input.

Output

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

Model Architecture

Conformer-CTC model is a non-autoregressive variant of Conformer model [1] for Automatic Speech Recognition which uses CTC loss/decoding instead of Transducer. You may find more info on the detail of this model here: Conformer-CTC Model.

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 vocabulary we use contains 28 characters:

[' ', "'", 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z']

Rare symbols with diacritics were replaced during preprocessing.

The tokenizers for these models were built using the text transcripts of the train set with this script. For vocabulary of size 128 we restrict maximum subtoken length to 2 symbols to avoid populating vocabulary with specific frequent words from the dataset. This does not affect the model performance and potentially helps to adapt to other domain without retraining tokenizer.

Full config can be found inside the .nemo files.

Datasets

All the models in this collection are trained on MCV-9.0 Kinyarwanda dataset, which contains around 2000 hours training, 32 hours of development and 32 hours of testing speech audios.

Performance

The list of the available models in this collection is shown in the following table. Performances of the ASR models are reported in terms of Word Error Rate (WER%) with greedy decoding.

Version Tokenizer Vocabulary Size Dev WER Test WER Train Dataset
1.11.0 SentencePiece BPE, maxlen=2 128 15.16 18.22 MCV-9.0 Train set

Limitations

Since this model was trained on publicly available speech datasets, the performance of this model might degrade for speech which includes technical terms, or vernacular that the model has not been trained on. The model might also perform worse for accented speech.

Deployment with NVIDIA Riva

For the best real-time accuracy, latency, and throughput, deploy the model with NVIDIA Riva, an accelerated speech AI SDK deployable on-prem, in all clouds, multi-cloud, hybrid, at the edge, and embedded. Additionally, Riva provides:

References