Speech-to-Text server

This is a key feature in the voice interaction systems.
Now, there are many online services with this functionality, but they are not applicable everywhere because they lead to security issues (data leaks), may brings delays in interaction (slow response), and so on.
When I designing systems that require voice interaction (automatic calls processing, voice request processing, etc), a frequent requirement is to use a local Speech-to-Text systems.
And this solution was written to resolve this problems.

The server was written in C, based on wstk, alphacephei framework and its models.
Capable to work on the regular servers, produces fast responses that suitable to build realtime dialog systems.

This is a commercial product, if you are interested in purchasing or have some questions, please visit a contact page.
There is an evaluation period with installation on your servers (preferred Ubuntu 22.04 x64).

Basic features:

• Neural based, complete locally system

  Allows to keep your personal data in safe

• Multilingual support

  There are open models for various languages

• Fine-tuning of models if necessary

  There are tools for it

• Speech contex definition

  Allows you to defined (in request) a dictionary of available words

• Speaker identify

  Allows to generate vector for speaker identification

• Capable to work on regular servers

  You don't need to purchase or rent some expensive hardware

• Supports Freeswitch

  There's a module for integration with Freeswitch (see mod_sivr_asr)
  Allows to use asr-api from dialplan or scripts

• Caching and preload models

  Allows to save memory and improve performance by sharing models

• Provides simple web api

  This allows to integrate STT service in various application (see example below)

• Supports formats: wav, mp3

  With various samplerate and channels.

• Supports os: Linux

--- Examples ---

Example #1 (simple request)

Request:
curl http://127.0.0.1:8801/v1/transcriptions -X POST -H "Authorization: Bearer secret" -H "Content-Type: multipart/form-data" -F language="en" -F smodel="small" -F file="@test.mp3"

Response (json):
{
 "text" : "hello world"
 }
        

Example #2 (with speakes identify)

Request:
curl http://127.0.0.1:8801/v1/transcriptions -X POST -H "Authorization: Bearer secret" -H "Content-Type: multipart/form-data" -F language="en" -F smodel="small" -F vmodel="default" -F file="@test.mp3"

Response (json):
{
 "spk" : [-0.644623, 1.023342, 2.575434, 0.623447, -0.602342, 1.0234234 -1.4824234 -0.021242, 0.824297, -0.152424, ... ],
 "spk_frames" : 81,
 "text" : "hello world"
}