» Current | 2020 | 2019 | 2018 | Subscribe

Linley Newsletter

Alibaba’s Ouroboros Is a Fast Talker

August 27, 2019

Author: Mike Demler

To provide driving directions, read an e-book, or perform other interactive functions, a digital assistant must be able to speak the user’s language fluently. Although Alexa, Cortana, Siri, and their counterparts may sound human, they all get their voices from text-to-speech (TTS) synthesis. Smartphones, smart speakers, and other consumer electronics with natural-language user interfaces (NLUIs) run TTS in the cloud, but Alibaba’s new Ouroboros speech synthesizer aims to move TTS to embedded devices.

The company’s data centers have deployed an Ouroboros FPGA prototype running on Xilinx VU9P boards. It now plans to integrate that architecture in a complete TTS chip. Alibaba says that in simulations of a 16nm ASIC, Ouroboros can synthesize speech with 30ms latency, matching the performance of the WaveNet model in Google data centers.

The simulations also demonstrate that Ouroboros can reduce the voice-generation component of a TTS system to 3W for a high-performance model or to as little as 200mW in a lower-performance embedded design. TTS, however, is a multistage process. The company plans to adopt in its ASIC the feature-extraction and mel-spectrogram-estimation components that DeepMind developed for Google. We expect those components will increase power consumption to at least 10W.

The 200mW synthesizer generates speech at 2x real time, or 10x slower than the high-performance model. That speed is adequate for low-cost IoT devices, toys, and other embedded systems. But having just 256 MAC units, the embedded Ouroboros is unlikely to match DeepMind’s voice quality, which in blind testing scores within 3% of digitized human speech.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article

Events

Linley Spring Processor Conference 2020
Coming April 7-8, 2020
Hyatt Regency, Santa Clara, CA
Register Now!
Linley Fall Processor Conference 2020
Coming October 28-29, 2020
Hyatt Regency, Santa Clara, CA
More Events »

Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »