sber-salute-speech-recognition

A library for getting audio transcriptions from the SBER Salute Speech service https://developers.sber.ru/docs/ru/salutespeech/recognition/recognition-sync.

Getting started

1. Create an account on https://developers.sber.ru/portal/products/smartspeech
2. Generate a client secret
3. Capture the auth key

Install

npm install sber-salute-speech-recognition

Usage

import {
  SberSaluteSpeechRecognitionService,
  AudioEncoding,
} from 'sber-salute-speech-recognition';

const recognitionService = new SberSaluteSpeechRecognitionService(AUTH_KEY);
const { text, normalizedText } = await recognitionService.speechToText(
  pathToAudioFile,
  AudioEncoding.MP3,
);

AccessToken Scope

The scope property is set during the instantiation of the SberSaluteSpeechRecognitionService class. If no value is provided, it defaults to Scope.Personal.

Here is an example of how to use the scope property:

import {
  SberSaluteSpeechRecognitionService,
  Scope,
} from 'sber-salute-speech-recognition';

const service = new SberSaluteSpeechRecognitionService(
  AUTH_KEY,
  undefined,
  Scope.Corporate
);

In this example, the scope property is set to Scope.Corporate what equals SALUTE_SPEECH_CORP. If you want to use the SALUTE_SPEECH_PERS scope, you can either pass Scope.Personal as the third argument or omit the third argument entirely, as it defaults to Scope.Personal.

Hints param

For improve speech recognition you can pass in speechToText method hints param

Here is example of how to use the hints param:

import {
  SberSaluteSpeechRecognitionService,
  AudioEncoding,
} from 'sber-salute-speech-recognition';

const recognitionService = new SberSaluteSpeechRecognitionService(AUTH_KEY);
const { text, normalizedText } = await recognitionService.speechToText(
  pathToAudioFile,
  AudioEncoding.MP3,
  {
    words: ['card', 'name'],
    enable_letters: true,
    eou_timeout: "2s"
  }
);

In this example , we pass object with props words, enable_letters, eou_timeout.

words - A list of words or phrases whose recognition we want to strengthen. Here you can list the words that the user is likely to pronounce

enable_letters - A short phrase model that improves recognition of single letters and short words. Possible values: true and false

eou_timeout - Setting up recognition of the end of a phrase (End of Utterance - eou). Such recognition will be expected after the end of the phrase for as many seconds as set in this parameter. Possible values are from 0.5 to 5 seconds.

Channels count

The channels_count allows to override the number of channels to recognize in the audio file. If not specified, the number of channels will be determined automatically.

Possible values: undefined - channels count will be determined automatically 1 - mono 2 - stereo

import {
  SberSaluteSpeechRecognitionService,
  AudioEncoding,
} from 'sber-salute-speech-recognition';

const recognitionService = new SberSaluteSpeechRecognitionService(AUTH_KEY);
const { text, normalizedText } = await recognitionService.speechToText(
  pathToAudioFile,
  AudioEncoding.MP3,
  1
);

In this example we set channels count to 1.