Duration: 1:15:36

8,590 views

147

In the popular field of text to speech, the goal is to transform the written or printed word into speech that is natural and intelligible. Today, the technology is being used in products and services to help people who are blind or have low vision consume digital content, power personal digital assistants that sound more realistic, and make it easier to do two things at once, such as listening to an article online while washing dishes, among other applications. Although the quality of synthesized speech has gotten better thanks to neural network-based end-to-end TTS, advancing neural TTS and allowing it to be more easily integrated into product development and deployment requires overcoming a variety of remaining challenges.

In this webinar, Senior Researcher Xu Tan will talk about these challenges, specifically the high computational cost and slow inference speed in online serving; word skipping and repeating issues, poor voice quality, and lack of voice controllability; the large amounts of training data needed for improved voice synthesis; and the practical challenges in TTS voice adaptation. He’ll introduce his team’s work addressing these challenges—including fast TTS, end-to-end TTS, low-resource TTS, and adaptive TTS—as well as discuss other critical questions and opportunities to pursue in the space.

Together, you'll explore:

■ An overview of text to speech, including its evolution
■ The important challenges in neural text to speech and how to address them with dedicated research
■ How to factor product development into your research

𝗥𝗲𝘀𝗼𝘂𝗿𝗰𝗲 𝗹𝗶𝘀𝘁:

■ Text to Speech (project page): https://www.microsoft.com/en-us/research/project/text-to-speech/
■ Xu Tan (publications page): https://www.microsoft.com/en-us/research/people/xuta/publications/
■ Speech Research Repository Master List (GitHub): https://speechresearch.github.io/
■ FastSpeech: Fast, Robust and Controllable Text to Speech (GitHub): https://speechresearch.github.io/fastspeech/
■ FastSpeech 2: Fast and High-Quality End-to-End Text to Speech (GitHub): https://speechresearch.github.io/fastspeech2/
■ AdaSpeech: Adaptive Text to Speech for Custom Voice (GitHub): https://speechresearch.github.io/adaspeech/
■ AdaSpeech 2: Adaptive Text to Speech with Untranscribed Data (GitHub): https://speechresearch.github.io/adaspeech2/
■ LightSpeech: Lightweight and Fast Text to Speech with Neural Architecture Search (GitHub): https://speechresearch.github.io/lightspeech/
■ LRSpeech: Extremely Low-Resource Speech Synthesis and Recognition (Github): https://speechresearch.github.io/lrspeech/
■ Neural Text-to-Speech previews five new languages with innovative models in the low-resource setting (blog): https://techcommunity.microsoft.com/t5/azure-ai/neural-text-to-speech-previews-five-new-languages-with/ba-p/1907604
■ Microsoft Azure Text to Speech: https://azure.microsoft.com/en-us/services/cognitive-services/text-to-speech
■ Microsoft Azure Custom Voice: https://speech.microsoft.com/customvoice
■ Xu Tan (Researcher profile): https://www.microsoft.com/en-us/research/people/xuta/

*This on-demand webinar features a previously recorded Q&A session and open captioning.

This webinar originally aired on May 20, 2021

Explore more Microsoft Research webinars: https://aka.ms/msrwebinars