GitHub - NVIDIA/NeMo-speech-data-processor: A toolkit for processing speech data and creating speech datasets
The Speech Data Processor (SDP) is a toolkit designed to simplify the processing of speech datasets. It minimizes the boilerplate code required and allows for easy sharing of processing steps. SDP's philosophy is to represent processing operations as 'processor' classes, which take in a path to a NeMo-style data manifest as input (or a path to the raw data directory if you do not have a NeMo-style manifest to start with), apply some processing to it, and then save the output manifest file.
Features
Creating Manifests: Generate manifests for your datasets.
Running ASR Inference: Automatically run ASR inference to remove utterances where the reference text differs greatly from ASR predictions.
Text Transformations: Apply text-based transformations to lines in the manifest.
Removing Inaccurate Transcripts: Remove lines from the manifest which may contain inaccurate transcripts.
Custom Processors: Write your own processor classes if the provided ones do not meet your needs.
Installation
SDP is officially supported for Python 3.10, but might work for other versions.
Clone the repository:
git clone https://github.com/NVIDIA/NeMo-speech-data-processor.git cd NeMo-speech-data-processorInstall dependencies:
pip install -r requirements/main.txtOptional: If you need to use ASR, NLP parts, or NeMo Text Processing, follow the NeMo installation instructions:
Example:
In this example we will load librispeech using SDP. For downloading all available data - replace config.yaml with all.yaml
For mini dataset - replace with mini.yaml.
python NeMo-speech-data-processor/main.py \ --config-path="dataset_configs/english/librispeech" \ --config-name="config.yaml" \ processors_to_run="0:" \ workspace_dir="librispeech_data_dir"Usage
Create a Configuration YAML File:
Here is a simplified example of a config.yaml file:
processors: - _target_: sdp.processors.CreateInitialManifestMCVoutput_manifest_file: "${data_split}_initial_manifest.json"language_id: es - _target_: sdp.processors.ASRInferencepretrained_model: "stt_es_quartznet15x5" - _target_: sdp.processors.SubRegexregex_params_list: - {"pattern": "¡", "repl": "."} - {"pattern": "ó", "repl": "o"}test_cases: - {input: {text: "hey!"}, output: {text: "hey."}} - _target_: sdp.processors.DropNonAlphabetalphabet: "abcdefghijklmnopqrstuvwxyzáéiñóúüABCDEFGHIJKLMNOPQRSTUVWXYZÁÉÍÑÓÚÜ"test_cases: - {input: {text: "test Тест ¡"}, output: null} - {input: {text: "test"}, output: {text: "test"}} - _target_: sdp.processors.KeepOnlySpecifiedFieldsoutput_manifest_file: "${data_split}_final_manifest.json"fields_to_keep: - "audio_filepath" - "text" - "duration"
Run the Processor:
Use the following command to process your dataset:
python <SDP_ROOT>/main.py \ --config-path="dataset_configs/<lang>/<dataset>/" \ --config-name="config.yaml" \ processors_to_run="all" \ data_split="train" \ workspace_dir="<dir_to_store_processed_data>"
To learn more about SDP, have a look at our
.
Contributing
We welcome community contributions! Please refer to the
for the process.