Deovrat's picture
Push model using huggingface_hub.
6e2cb37 verified
|
raw
history blame
2.61 kB
metadata
{}
## IndicConformer

IndicConformer is an Hybrid RNNT conformer model built for Konkani.

## AI4Bharat NeMo:

To load, train, fine-tune or play with the model you will need to install [AI4Bharat NeMo](https://github.com/AI4Bharat/NeMo). We recommend you install it using the command shown below
```
git clone https://github.com/AI4Bharat/NeMo.git && cd NeMo && git checkout nemo-v2 && bash reinstall.sh
```

## Usage

```bash
$ python inference.py --help
usage: inference.py [-h] -c CHECKPOINT -f AUDIO_FILEPATH -d (cpu,cuda) -l LANGUAGE_CODE

options:
-h, --help            show this help message and exit
-c CHECKPOINT, --checkpoint CHECKPOINT
                        Path to .nemo file
-f AUDIO_FILEPATH, --audio_filepath AUDIO_FILEPATH
                        Audio filepath
-d (cpu,cuda), --device (cpu,cuda)
                        Device (cpu/gpu)
-l LANGUAGE_CODE, --language_code LANGUAGE_CODE
                        Language Code (eg. hi)
```

## Example command
```
python inference.py -c ai4b_indicConformer_hi.nemo -f hindi-16khz.wav -d cuda -l hi
```
Expected output - 

```
Loading model..
...
Transcibing..
----------
Transcript: 
Took ** seconds.
----------
```

### Input

This model accepts 16000 KHz 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 is a onformer-Large model, consisting of 120M parameters, as the encoder, with a hybrid CTC-RNNT  decoder. The model has 17 conformer blocks with
512 as the model dimension.

## Training

<ADD INFORMATION ABOUT HOW THE MODEL WAS TRAINED - HOW MANY EPOCHS, AMOUNT OF COMPUTE ETC>

### Datasets

<LIST THE NAME AND SPLITS OF DATASETS USED TO TRAIN THIS MODEL (ALONG WITH LANGUAGE AND ANY ADDITIONAL INFORMATION)>

## Performance

<LIST THE SCORES OF THE MODEL -
    OR
USE THE Hugging Face Evaluate LiBRARY TO UPLOAD METRICS>

## Limitations

<DECLARE ANY POTENTIAL LIMITATIONS OF THE MODEL>

Eg:
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.


## References

<ADD ANY REFERENCES HERE AS NEEDED>

[1] [AI4Bharat NeMo Toolkit](https://github.com/AI4Bharat/NeMo)