ypluit's picture
Update README.md
df70866
metadata
language:
  - kr
license: cc-by-4.0
library_name: nemo
datasets:
  - RealCallData
thumbnail: null
tags:
  - automatic-speech-recognition
  - speech
  - audio
  - Citrinet1024
  - NeMo
  - pytorch
model-index:
  - name: stt_kr_citrinet1024_PublicCallCenter_1000H_0.22
    results: []

Model Overview

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 [1], 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.ASRModel.from_pretrained("ypluit/stt_kr_citrinet1024_PublicCallCenter_1000H_0.22")

Transcribing using Python

First, let's get a sample

get any korean telephone voice wave file

Then simply do:

asr_model.transcribe(['sample-kr.wav'])

Transcribing many audio files

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

Input

This model accepts 16000Hz Mono-channel Audio (wav files) as input.

Output

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

Model Architecture

See nemo toolkit and reference papers.

Training

Learned about 30 days on 2 A6000

Datasets

Private call center real data (1100hour)

Performance

< 0.13 CER

Limitations

This model was trained with 650 hours of Korean telephone voice data for customer service in a call center. might be Poor performance for general-purpose dialogue and specific accents.

References

[1] NVIDIA NeMo Toolkit