Edit model card

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("eesungkim/stt_kr_conformer_transducer_large")

Transcribing using Python

First, let's get a sample

wget https://dldata-public.s3.us-east-2.amazonaws.com/sample-kor.wav

Then simply do:


Transcribing many audio files

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


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


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

Model Architecture

Conformer-Transducer model is an autoregressive variant of Conformer model [2] for Automatic Speech Recognition which uses Transducer loss/decoding. You may find more info on the detail of this model here: Conformer-Transducer Model.


The model was finetuned based on the pre-trained English Model for over several epochs.

There are several transcribing and sub-word modeling methods for Korean speech recognition. This model uses sentencepiece subwords of Hangul characters based on phonetic transcription using Google Sentencepiece Tokenizer [3].


All the models in this collection are trained on Ksponspeech dataset, which is an open-domain dialog corpus recorded by 2,000 native Korean speakers in a controlled and quiet environment. The standard split dataset consists of 965 hours of training set, 4 hours of development set, 3 hours of test-clean, and 4 hours of test-other.


Version Tokenizer eval_clean CER eval_other CER eval_clean WER eval_other WER
v1.7.0rc SentencePiece Char 6.94% 7.38% 19.49% 22.73%


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

This model produces a spoken-form token sequence. If you want to have a written form, you can consider applying inverse text normalization.


[1] NVIDIA NeMo Toolkit

[2] Conformer: Convolution-augmented Transformer for Speech Recognition

[3] Google Sentencepiece Tokenizer

Downloads last month
Hosted inference API
This model can be loaded on the Inference API on-demand.