8khz telephony audio ASR

[utkarshpandey12]

Hi,

Just wanted to know if Paddlespeech can detect speech from .wav audio files sampled at 8khz. If yes do I just staright away pass the audio to using the CLI method or do I need to edit sample rate somewhere as pass it as a partameter?

What is the correct approach to use PaddleSpeech for 8khz sampled .wav audio files?

Best
Utkarsh

[utkarshpandey12]

Please suggest how to get started with transcribing 8khz .wav audio files using Python API or CLI option.
Do I need to train the model from scratch as I see all the models have been trained using 16khz Audio data?.
As explained above I want to transcribe 8khz files either by using CLI/Python API.

When I am submitting 8khz file using the below CLI command it is returning me empty string however the audio file contains speech
( test) $paddlespeech asr --model transformer_librispeech --lang en --input test.wav

I also tried giving the --sample_rate parameter = 8000 using the below command still it return empty string
( test) $paddlespeech asr --model transformer_librispeech --lang en --input test.wav --sample_rate 8000

@ryanrussell @harisankarh @ZeyuChen @wangkuiyi

[yt605155624]

we only support 16k ASR model till now, maybe you should train your own 8K ASR model or upsample your 8k wavs to 16k first and use our 16K models

[utkarshpandey12]

So is this flow of steps correct to finish with the training process?
Please correct me if I am missing something.

git clone
cd examples
cd callcenter
cd asr1
~/asr1
~/asr1 source ${MAIN_ROOT}/utils/parse_options.sh
~/asr1 bash run.sh --stage 0 --stop_stage 6

[yt605155624]

Almost, but I suggest you run step by step to ensure that you know the function of each step

bash run.sh --stage 0 --stop_stage 0
bash run.sh --stage 1 --stop_stage 1
...

[utkarshpandey12]

I started the 1st step in training the callcenter dataset using the command below
bash run.sh --stage 0 --stop_stage 0

I'm getting following errors. Can you please tell me what am I missing?

center/asr1$ bash run.sh --stage 0 --stop_stage 0
checkpoint name conformer
mv: cannot stat 'data/manifest.train': No such file or directory
mv: cannot stat 'data/manifest.dev': No such file or directory
mv: cannot stat 'data/manifest.test': No such file or directory
----------- compute_mean_std.py Arguments -----------
delta_delta: 0
feat_dim: 80
manifest_path: data/manifest.train.raw
num_samples: -1
num_workers: 4
output_path: data/mean_std.json
sample_rate: 8000
spectrum_type: fbank
stride_ms: 10
target_dB: -20
use_dB_normalization: 0
window_ms: 25

2022-07-13 19:43:51.704 | INFO | paddlespeech.s2t.frontend.augmentor.augmentation:init:122 - Augmentation: []
Traceback (most recent call last):
File "/home/utkarsh/Desktop/peddle/PaddleSpeech/utils/compute_mean_std.py", line 87, in
main()
File "/home/utkarsh/Desktop/peddle/PaddleSpeech/utils/compute_mean_std.py", line 77, in main
normalizer = FeatureNormalizer(
File "/home/utkarsh/Desktop/peddle/PaddleSpeech/paddlespeech/s2t/frontend/normalizer.py", line 117, in init
self._compute_mean_std(manifest_path, featurize_func, num_samples,
File "/home/utkarsh/Desktop/peddle/PaddleSpeech/paddlespeech/s2t/frontend/normalizer.py", line 166, in _compute_mean_std
dataset = AudioDataset(manifest_path, num_samples, self._rng)
File "/home/utkarsh/Desktop/peddle/PaddleSpeech/paddlespeech/s2t/frontend/normalizer.py", line 65, in init
with jsonlines.open(manifest_path, 'r') as reader:
File "/home/utkarsh/.pyenv/versions/peddle/lib/python3.9/site-packages/jsonlines/jsonlines.py", line 627, in open
fp = builtins.open(file, mode=mode + "t", encoding=encoding)
FileNotFoundError: [Errno 2] No such file or directory: 'data/manifest.train.raw'
Compute mean and stddev failed. Terminated.

[zh794390558]

You need to prepare 8k train data yourself, and generate manifest for train,dev,test. You can see this as reference. Callcenter egs only has receipt to train, since we do not have 8k opensource data. Or you can using Aishell or Librispeech egs, change wav to 8k sample rate, then using model/preprocess config in Callcenter egs.

[utkarshpandey12]

Just want to confirm on one point

Can I have training examples where there is no audio and for those audio files I will give an empty string in the dataset in the transcript txt file.
Or will it cause some issue ?