Skip to content

Persian STT v0.1.0

Latest
Compare
Choose a tag to compare
@JRMeyer JRMeyer released this 11 Jul 18:14
· 14 commits to main since this release
b14378c

Persian STT v0.1.0

Jump to section:

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} }
  • persian-tts GitHub Repo
  • 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%
  • More about the model at persian-tts repo

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: .65

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) model.tflite 45.3M
Language model kenlm.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.