4 ms·

How many business answer the phone by saying "How can I help you?", are they lying? Are they not even close to have a system like that?

by philca 8y ago

jacksmith21006 8y ago
Google shared they removed personal information from the calls like the names of the locations. They already have the TTS aspect out in beta. It uses something called WaveNet. Here this might help. https://deepmind.com/blog/wavenet-generative-model-raw-audio/ https://deepmind.com/blog/wavenet-generative-model-raw-audio... and "Introducing Cloud Text-to-Speech powered by DeepMind WaveNet technology" https://cloudplatform.googleblog.com/2018/03/introducing-Cloud-Text-to-Speech-powered-by-Deepmind-WaveNet-technology.html https://cloudplatform.googleblog.com/2018/03/introducing-Clo... Testing will start on the rest later this summer. https://ai.googleblog.com/2018/05/duplex-ai-system-for-natural-conversation.html https://ai.googleblog.com/2018/05/duplex-ai-system-for-natur... On the link some more examples of calls. Pretty cool stuff. Really just shows how far ahead Google is in the AI space at every layer of the stack. The TTS at scale required the TPUs. Doing 16k cycles through a DNN in real-time is a really intensive compute task but then to offer at a competitive price to using the old way needs custom silicon.
- philca 8y ago
  I missed https://news.ycombinator.com/item?id=17094231 https://news.ycombinator.com/item?id=17094231 Ah I was not the only one finding it suspicious, Thank you for your links, pretty interesting research quite far away from a real product though. This Weekend was the bay area makerfaire, where the Google stand was in front of the one of Nintendo, they were showing the exact same tech, but the Google one was really underwhelming, it was actually interesting to compare it. I don't trust Google, ok let's wait for a product I guess.
  - jacksmith21006 8y ago
    The voice part is already in beta and on the Google Home. https://cloud.google.com/text-to-speech/docs/wavenet https://cloud.google.com/text-to-speech/docs/wavenet That is the part that freaked out people as you can not tell the difference between humans and computers any longer. Google is using a DNN at 16k cycles a second in real-time. Suspect that was only possible to do at scale and at a reasonable cost with the TPUs. They are competing with a much less compute intensive approach. But Wavenet gets a much better result. The other aspects of Duplex will enter testing later this summer. So not long to wait. "Google Duplex: An AI System for Accomplishing Real-World Tasks Over the Phone" https://ai.googleblog.com/2018/05/duplex-ai-system-for-natural-conversation.html https://ai.googleblog.com/2018/05/duplex-ai-system-for-natur... This is the type of thing that contributed me buying the Google Home. Google is just so far ahead of everyone else in AI. Really every layer of the stack.

How many business answer the phone by saying "How can I help you?", are they lying? Are they not even close to have a system like that?

by philca 8y ago

How many business answer the phone by saying "How can I help you?", are they lying? Are they not even close to have a system like that?

jacksmith21006 8y ago

Google shared they removed personal information from the calls like the names of the locations. They already have the TTS aspect out in beta. It uses something called WaveNet. Here this might help. https://deepmind.com/blog/wavenet-generative-model-raw-audio/ https://deepmind.com/blog/wavenet-generative-model-raw-audio... and "Introducing Cloud Text-to-Speech powered by DeepMind WaveNet technology" https://cloudplatform.googleblog.com/2018/03/introducing-Cloud-Text-to-Speech-powered-by-Deepmind-WaveNet-technology.html https://cloudplatform.googleblog.com/2018/03/introducing-Clo... Testing will start on the rest later this summer. https://ai.googleblog.com/2018/05/duplex-ai-system-for-natural-conversation.html https://ai.googleblog.com/2018/05/duplex-ai-system-for-natur... On the link some more examples of calls. Pretty cool stuff. Really just shows how far ahead Google is in the AI space at every layer of the stack. The TTS at scale required the TPUs. Doing 16k cycles through a DNN in real-time is a really intensive compute task but then to offer at a competitive price to using the old way needs custom silicon.

philca 8y ago
I missed https://news.ycombinator.com/item?id=17094231 https://news.ycombinator.com/item?id=17094231 Ah I was not the only one finding it suspicious, Thank you for your links, pretty interesting research quite far away from a real product though. This Weekend was the bay area makerfaire, where the Google stand was in front of the one of Nintendo, they were showing the exact same tech, but the Google one was really underwhelming, it was actually interesting to compare it. I don't trust Google, ok let's wait for a product I guess.
- jacksmith21006 8y ago
  The voice part is already in beta and on the Google Home. https://cloud.google.com/text-to-speech/docs/wavenet https://cloud.google.com/text-to-speech/docs/wavenet That is the part that freaked out people as you can not tell the difference between humans and computers any longer. Google is using a DNN at 16k cycles a second in real-time. Suspect that was only possible to do at scale and at a reasonable cost with the TPUs. They are competing with a much less compute intensive approach. But Wavenet gets a much better result. The other aspects of Duplex will enter testing later this summer. So not long to wait. "Google Duplex: An AI System for Accomplishing Real-World Tasks Over the Phone" https://ai.googleblog.com/2018/05/duplex-ai-system-for-natural-conversation.html https://ai.googleblog.com/2018/05/duplex-ai-system-for-natur... This is the type of thing that contributed me buying the Google Home. Google is just so far ahead of everyone else in AI. Really every layer of the stack.