Skip to content

English STT v1.0.0-digits

Compare
Choose a tag to compare
@JRMeyer JRMeyer released this 03 Oct 19:27
· 40 commits to main since this release

English STT v1.0.0 (digits)

Jump to section:

Model details

  • Person or organization developing model: Maintained by Coqui.
  • Model language: English / English / en
  • Model date: October 3, 2021
  • Model type: Small vocabulary Speech-to-Text
  • Model version: v1.0.0-digits
  • Compatible with 🐸 STT version: v1.0.0
  • License: Apache 2.0
  • Citation details: @techreport{english-stt, author = {Coqui}, title = {English STT v1.0.0}, institution = {Coqui}, address = {\url{https://coqui.ai/models}} year = {2021}, month = {October}, number = {STT-EN-1.0.0} }
  • 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

Closed vocabulary (digits "zero" through "nine") Speech-to-Text for the English Language on 16kHz, mono-channel audio. This acoustic model and language model pair will only be able to recognize the words {"zero","one","two","three","four","five","six","seven","eight" and "nine"}, which is a common use case in IVR systems.

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

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. The acoustic model comes in two forms: quantized and unquantized. There is a size<->accuracy trade-off for acoustic model quantization. For this combination of acoustic model and language model, we optimize for small size.

Model type Vocabulary Filename Size
Acoustic model open model_quantized.tflite 46M
Language model small digits.scorer 1.7K

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 7.0 English (custom Coqui train/dev/test splits), LibriSpeech, and Multilingual Librispeech. In total approximately ~47,000 hours of data.

Evaluation data

The validation ("dev") sets came from CV, Librispeech, and MLS. Testing accuracy is reported for MLS and Librispeech.

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.