Skip to main content

Aimyvoice TTS

Aimyvoice is a speech synthesis model from Just AI. The model is distinguished by:

  • high processing speed
  • a streaming mode for generating long texts
  • support for SSML for tweaking pronunciation
  • the ability to create derived voices with a preset pronunciation dictionary

Aimyvoice uses Caila as a hosting platform, and the model is available:

Voice testing​

Voices can be tested on the service page.

  You can select a voice from a list or set synthesis parameters in JSON format.
</p>

Connect using API​

The speech synthesis service can be accessed in a standard way for all Caila services.

In addition, there are two specialised adapters for TTS services.

Caila API​

HTTP TTS​

  • POST https://app.caila.io/api/mlpgate/account/just-ai/model/aimyvoice-facade/tts. In this method, all parameters, including voice name and encoding parameters, must be passed in the request body. The response is returned as application/octet-stream in LINEAR16_PCM encoding (and not JSON).

Request example:

POST https://app.caila.io/api/mlpgate/account/just-ai/model/aimyvoice-facade/tts

{
  "text": "Hello",
  "voice": "Marry",
  "outputAudioSpec": {
    "audioEncoding": "LINEAR16_PCM",
    "sampleRateHertz": 8000,
    "chunkSizeKb": 1
  }
}

GRPC processSynthesis​

A method for performing synthesis via GRPC. This method accepts the same data as the previous two methods. The response returns a stream of chunks, each containing a small audio segment in PCM format.

rpc processSynthesis(ClientTtsRequestProto) returns (stream ClientTtsResponseProto)

To learn how to connect to Caila via GRPC, read this.

Pronunciation settings​

Aimyvoice supports a rich set of markup for audio control.

Try synthesising this text:

<break time="0.8s" breath="1" breath_power="0.8" breath_dura="1.2"/>
My uncle <break time="1s"/> is of the most honest rules. <break time="250ms"/> When seriously unwell.
<break time="0.5s"/> He made others respect him. And you couldn’t think of anything better.

For more information on markup, see the original Aimyvoice documentation

Creating derived voices​

In Caila you can create a derived voice by configuring a pronunciation dictionary and some other voice parameters. Let’s look at how this can be done.

Create a dataset on the https://caila.io/workspace/dataset website with the json/tts-dictionary type.

Populate the dataset with rules.

Train the service on the https://caila.io/catalog/just-ai/aimyvoice-custom website with your rules.

Test the resulting derived service in My space.

  You can find the service on the Services/Fitted tab.
</p>