Skip to content

Latest commit

 

History

History
153 lines (116 loc) · 7.63 KB

README.md

File metadata and controls

153 lines (116 loc) · 7.63 KB
Raw

Persian (فارسی) STT Model

This is a Persian DeepSpeech model that is currently in development. It is licensed under GNU GPLv3.

Jump to section:

Data Preprocessing

Common-Voice

  • persian_stt.ipynb contains a notebook that demonstrates how to preprocess the data.
  • The Common-Voice dataset was obtained from here.
  • Using STT's bin/import_cv2.py script, the data was matched to the alphabet converted to CSV files.
  • The generated validated.csv was passed to auto_input_dataset=metadata in the configuation step.

Text Corpus

  • The raw text corpus was obtained from here.
  • Then, the corpus was normalized, tokenized, and cleaned using the persianify function and _HTMLStripper in the notebook.

Training

  • The model was trained 2 separate times with force_initialize_learning_rate. learning_rate was increased for the second run from 0.00012 to 0.0004.
  • Training was stopped at 1896648 steps after optimizing the model.

Initial Configuration

  • Tranfer learning from the STT English model was enabled (with drop_source_layers=2)
  • The model was trained with the following configuration:
initialize_globals_from_args(
    scorer_path='/content/kenlm-persian.scorer',
    dropout_rate=0.175,
    learning_rate=0.000095,
    force_initialize_learning_rate=True,
    train_cudnn=True,
    reduce_lr_on_plateau=True,
    plateau_epochs=3,
    plateau_reduction=0.2,
    auto_input_dataset=metadata,
    checkpoint_dir=checkpoint_dir,
    export_dir=export_dir,
    epochs=200,
    train_batch_size=128,
    dev_batch_size=128,
    test_batch_size=384,
)
  • Then, the dropout rate was decreased as the learning rate was decreasing due to the plateau to reach the minimum loss values.
  • Final test: WER: 0.108172, CER: 0.025068, loss: 1.292882
  • It is worth noting that many of the errors are due to the language syntax. For example, the model is unable to distinguish between the following words by themselves:
    امیررضا and امیر رضا آهودخت and آهو دخت امیرآرشام and امیر آرشام.
  • Thankfully, the scorer is able to handle most of these cases, according to the context.

Release

  • The acoustic and language models are available in releases

Model card for Persian STT v0.1.0

Model details

  • Person or organization developing model: Maintained by oct4pie.
  • Model language: Persian / Farsi / fa, fa-IR
  • Model date: June 21, 2022
  • Model type: Speech-to-Text
  • Model version: v0.1.0
  • Compatible with 🐸 STT version: v1.3.0
  • License: GNU Lesser General Public License v3.0
  • Citation details: @techreport{persian-stt, author = {Mehdi Hajmollaahmad Naraghi}, title = {Persian STT v0.1.0}, institution = {Coqui}, address = {\url{https://coqui.ai/models}} year = {2022}, month = {June}, number = {STT-FA-0.1.0} }
  • Where to send questions or comments about the model: You can leave an issue on STT issues, open a new discussion on STT discussions, or chat with us on Gitter.

Intended use

Speech-to-Text for the Persian Language on 16kHz, mono-channel audio.

Performance Factors

Factors relevant to Speech-to-Text performance include but are not limited to speaker demographics, recording quality, and background noise. Read more about STT performance factors here.

Metrics

STT models are usually evaluated in terms of their transcription accuracy, deployment Real-Time Factor, and model size on disk.

Transcription Accuracy

Using the language model with settings lm_alpha=0.36669178512950323 and lm_beta=0.3457913671678824 (found via lm_optimizer.py):

  • Common-Voice clean: WER: 10.81%, CER: 2.506%

Real-Time Factor

Real-Time Factor (RTF) is defined as proccesing-time / length-of-audio. The exact real-time factor of an STT model will depend on the hardware setup, so you may experience a different RTF.

Recorded average RTF on laptop CPU: .72

Model Size

For STT, you always must deploy an acoustic model, and it is often the case you also will want to deploy an application-specific language model.

Model type Filename Size
Acoustic model (tflite) persian_stt.tflite 45.3M
Acoustic model (pb) persian_stt.pb 181M
Acoustic model (pbmm) persian_stt.pbmm 181M
Language model kenlm-persian.scorer 1.63GB

Approaches to uncertainty and variability

Confidence scores and multiple paths from the decoding beam can be used to measure model uncertainty and provide multiple, variable transcripts for any processed audio.

Training data

This model was trained on the following corpora: Common Voice 9.0 Persian (cleaned and with custom train/dev/test splits). In total approximately ~271 hours of data.

Evaluation data

The validation ("dev") sets were cleaned and generated from Common Voice 9.0 Persian.

Ethical considerations

Deploying a Speech-to-Text model into any production setting has ethical implications. You should consider these implications before use.

Demographic Bias

You should assume every machine learning model has demographic bias unless proven otherwise. For STT models, it is often the case that transcription accuracy is better for men than it is for women. If you are using this model in production, you should acknowledge this as a potential issue.

Surveillance

Speech-to-Text may be mis-used to invade the privacy of others by recording and mining information from private conversations. This kind of individual privacy is protected by law in may countries. You should not assume consent to record and analyze private speech.

Caveats and recommendations

Machine learning models (like this STT model) perform best on data that is similar to the data on which they were trained. Read about what to expect from an STT model with regard to your data here.

In most applications, it is recommended that you train your own language model to improve transcription accuracy on your speech data.