As part of a strategy to partner with highly innovative technology companies, Graphcore is collaborating with Gridspace to drive innovation in conversational artificial intelligence (AI). Envisioning a contact center automation platform accelerated by intelligence processing units (IPUs), engineers at Graphcore and Gridspace are working together to explore IPU-based voice applications toward the goal of human parity.

In under two weeks, Graphcore, with support from Gridspace engineers, implemented a high-performance model, Transformer Transducer, to accelerate the training of over 700G/10k hours of speech with a state-of-the-art (SOTA) word error rate. The collaboration produced a 100M parameter speech application in a matter of days, rather than weeks, exploiting the linear scaling of up to 64 Graphcore IPU chips in the Bow Pod 64 system.

The Transformer Transducer model for speech recognition achieves SOTA performance on word error rate. The model consists of transformer layers to encode the audio, long short-term memory (LSTM) layers to encode text and a Joint Net to combine them. At the head is a transducer loss that is designed to align the audio and text, which can be of varying lengths.

This is one of many examples of the out-of-the-box performance of the Poplar SDK developed by Graphcore. In addition, Graphcore has released other off-the-shelf tooling and examples, which you can find easily and extend for different datasets and domain-specific tasks (e.g., earning call dataset, other language corpora).

As a result of the easy-to-use and easy-to-integrate examples, our time to value is improved by the speedup from the hardware and agility of development iterations.

We are looking forward to continuing our excellent collaboration with the innovators in machine intelligence in order to create the next breakthroughs in conversational artificial intelligence.

#ListentoGrace at www.gridspace.com.