DeepMind and Google use artificial intelligence to recreate the voice of former NFL running back Tim Shaw.

2019-12-19

In August of this year, Google AI researchers collaborated with the ALS Therapy Development Institute to share details of the Euphonia project, a speech-to-text transcription service for individuals with speech impairments. They demonstrated that by using audio datasets from native and non-native English speakers and the Parrotron technology (an AI tool for individuals with disabilities), the quality of speech synthesis and generation can be significantly improved. Recently, in a case study, Google researchers and Alphabet's DeepMind team employed Euphonia to attempt to reproduce the original voice of Tim Shaw, a former NFL football guard who played for the Carolina Panthers, Jacksonville Jaguars, Chicago Bears, and Tennessee Titans before retiring in 2013. Approximately six years ago, Shaw was diagnosed with amyotrophic lateral sclerosis (ALS), which left him unable to speak, swallow, or even breathe without assistance. Over a period of six months, the joint research team utilized a generative AI model called WaveNet to synthesize speech from audio samples taken before Shaw's diagnosis of ALS. WaveNet is capable of mimicking intonation and accentuation. Compared to previous speech generation models, the speech segments generated by WaveNet are more convincing. Google stated that, based on average opinion scores, it has reduced the quality gap between synthesized speech and human speech by 70%—and it is more efficient. When running on Google's Tensor Processing Units (TPUs), it only takes 50 milliseconds to create a 1-second speech sample. WaveNet has already been used to generate custom voices for Google's conversation platforms and Google Assistant. Recently, it has also been employed on the Google Cloud Platform to generate dozens of new voices and voice variants for Google's Cloud Text-to-Speech service—38 in August alone. Fine-tuning has proven to be crucial in achieving high-quality synthesis with minimal training data. In order to reproduce Shaw's voice, the Google and DeepMind teams adopted a method proposed in a research paper published last year ("Sample-Efficient Adaptive Text-to-Speech"), which involves pre-training a large-scale WaveNet model with thousands of speakers over a few days until it can produce basic speech with natural intonation. At this point, the model is fine-tuned with a small amount of data from the target speaker, giving the generated speech the characteristics of that speaker.

RECENT AI NEWS

RECENT AI TOOLS