Lately, Google has been focusing on the integration of Artificial Intelligence into its products, and the recent introduction of Translatotron is an example of it. It is an end-to-end and speech to speech translation model built by Google AI engineers.

Translatotron is capable of translating from one language to another through its single sequence to sequence AI model. The translation model magnificently translated two datasets from Spanish to English during its demonstration.

There are three components of this speech to speech translation system:

1. Speech Recognition: the speech can be converted into text form.
2. Machine Translation: already converted text can then be translated into another language.
3. Text-to-Speech Synthesis (TTS): the text that is translated can be converted into targeted language’s speech.

These systems are driven by many other speech-to-speech translation services like Google Translate.

In 2016, researchers for the first time presented the idea of practically developing a single sequence-to-sequence model which is capable of speech to text translation. This further led to an idea of building an end-to-end speech translation model. And finally, after three years, Google engineers introduced Translatotron.

By 2017, Google proved that such models can overtake the typical cascade models. Recently, there have been several new proposals aiming to improve the end to end speech to text translation models.

Translatotron, like cascaded system, does not require intermediate text representation in any of the languages. Source spectrograms are taken as input in its sequence-to-sequence network to generate the spectrogram of targeted language’s text.

This new sequence-to-sequence translation model is based on two different trained components:

• Neural Vocoder: output spectrogram is changed into time-domain waveforms.
• Speaker Encoder: the source speaker’s voice is kept in synthesized translated speech.

The quality of Translatotron was authorized by Google AI engineers using BLEU (Bilingual Evaluation Understudy) score. Results could have been lower than the typical cascade system but the team was able to show the effectiveness of end-to-end direct speech-to-speech translation.

The characteristics of real vocals are kept by the Translatotron with the addition of a speaker encoder network that makes speech sound as natural as possible. It is the first end to end model that is capable of directly translating from one language to another by keeping the original vocal sound. Engineers are taking it as an initial step towards future innovations on end-to-end speech-to-speech translation system.

Transtust is a reliable Vietanmese Translation Company that delivers each project with professionalism and accuracy . Our company is providing top quality services to our client, no less than AI Google Model.

Reference:

https://www.technotification.com/2019/05/google-ai-introduces-translatotron.html