Miss any of our Open RFC calls?Watch the recordings here! »

@bottlenose/rxtranscribe

0.7.0 • Public • Published

CircleCI License: MIT Version

Description

This package allows real-time speech-to-text (STT) functionality to be performed on audio streams. It offers numerous strategies for how the STT can be performed including the following pipelines: the opensource DeepSpeech architecture, the Amazon Transcribe API and Google Speech-to-text API.

Currently, audio can be passed in as a stream of Buffer objects containing audio data encoded using one of the following

Installation

yarn add @bottlenose/rxtranscribe
npm i --save @bottlenose/rxtranscribe

DeepSpeech

To run the DeepSpeech pipeline, download the DeepSpeech model, unzip it and pass the model directory to the toDeepSpeech operator like this: toDeepSpeech({modelDir: 'path/to/deepseech-models-0.7.0'}).

AWS Transcribe

To run the AWS Transcribe pipeline, you'll need a valid ACCESS_KEY_ID and SECRET_ACCESS_KEY with permissions to run AWS Transcribe.

GCP Speech-to-text

  • To run the GCP speech-to-text pipeline, you'll need a valid JSON file containing GCP credentials.
  • The project will need to have the speech-to-text API enabled.
  • You may need to set GOOGLE_APPLICATION_CREDENTIALS environment variable so that it contains the path of your credentials file.

💡 In the future, it would be nicer to have an npm library that can install Sox so that yarn install would handle all the dependencies.

Compatibility

Platform Support
node.js (>12)
Browsers
React Native
Electron

💡 This package has only been tested in the node.js environment. If it doesn't work isomorphically, it could probably be modified to do so without much effort. The authors haven't done so because running it on client devices does not seem like an advisable design for production situations. But if you want to take a stab at implementing isomorphic support, contact us!

Basic Usage

import {map} from 'rxjs/operators';
import {toDeepSpeech} from '@bottlenose/rxtranscribe';
 
// The pipeline takes a stream of .wav audio chunks (Buffer, String, Blob or Typed Array)
const buffer$ = pcmChunkEncodedAs16BitIntegers$.pipe(
  map(chunk => Buffer.from(chunk, 'base64')),
  toDeepSpeech({modelDir: '/path/to/deepspeech-models-0.7.0'})
);

Documentation & Guides

Install

npm i @bottlenose/rxtranscribe

DownloadsWeekly Downloads

11

Version

0.7.0

License

MIT

Unpacked Size

69.5 kB

Total Files

35

Last publish

Collaborators

  • avatar
  • avatar