4 ms·

I've done something similar with Sphinx, which worked quite well too. However, I feel the real challenge to productify such a technology would be to make this

I've done something similar with Sphinx, which worked quite well too. However, I feel the real challenge to productify such a technology would be to make this system recognise utterances spoken casually, over loud music, over friends talking in the background. Also, how would the speech to text recognise the user's commands over some sound playing from its own speaker. Suppose I set up a text-to-speech system to read a book to me. I want to be able to make it repeat a paragraph (because my attention had drifted away). However, the input channel of the microphone would contain not only me shouting "stop", but also the output of the text-to-speech system. We humans have a way of factoring out our own voice when we are listening to someone else. Is there a way to make machines do that?

rcorcs 12y ago
It is doable with some analysis of the sound signal. Honda's ASIMO does that, somehow. But I think it would be a current scientific research topic.
- sherjilozair 12y ago
  Any links on relevant material? I'm okay for delving into current research topic.
  - rcorcs 12y ago
    These are very interesting papers from Google: http://research.google.com/pubs/SpeechProcessing.html http://research.google.com/pubs/SpeechProcessing.html
srcreigh 12y ago
(These are my humbly presented thoughts, I am certainly no expert) The crux of any method of differentiating the user's signal from background noise, I imagine, would look to leverage data that has a higher signal-to-noise ratio than the naive microphone recording of the environment. At a high level this is how I would expect we humans do this. We have some sort of sensory data from us feeling the vibrations of our voice, and use this to "factor out" our own voice from what our ears perceive. In terms of how this might work to factor background noise out of a microphone recording, if we had some kind of hardware (phone, smartwatch, keyboard) that is able to sense the vibrations of the user's voice (via phone in the pocket, watch on the wrist, keyboard that you're resting your palms on as you type), we might be able to use some signal processing algorithms with that signal and the environmental noise to factor out the user's voice. (One might ask "well why not just use the signal from the hardware then?". In terms of quality, I'm not sure if it would be enough---intuitively, I'm guessing it might sound muffled if you played back the recordings of your voice as picked up by your palms resting on a keyboard. But, it might be good enough data that we can use some signal processing algorithm to tune our noisey, but high-quality, data for the user's voice.)

rcorcs 12y ago

It is doable with some analysis of the sound signal. Honda's ASIMO does that, somehow. But I think it would be a current scientific research topic.

sherjilozair 12y ago
Any links on relevant material? I'm okay for delving into current research topic.
- rcorcs 12y ago
  These are very interesting papers from Google: http://research.google.com/pubs/SpeechProcessing.html http://research.google.com/pubs/SpeechProcessing.html

srcreigh 12y ago

(These are my humbly presented thoughts, I am certainly no expert) The crux of any method of differentiating the user's signal from background noise, I imagine, would look to leverage data that has a higher signal-to-noise ratio than the naive microphone recording of the environment. At a high level this is how I would expect we humans do this. We have some sort of sensory data from us feeling the vibrations of our voice, and use this to "factor out" our own voice from what our ears perceive. In terms of how this might work to factor background noise out of a microphone recording, if we had some kind of hardware (phone, smartwatch, keyboard) that is able to sense the vibrations of the user's voice (via phone in the pocket, watch on the wrist, keyboard that you're resting your palms on as you type), we might be able to use some signal processing algorithms with that signal and the environmental noise to factor out the user's voice. (One might ask "well why not just use the signal from the hardware then?". In terms of quality, I'm not sure if it would be enough---intuitively, I'm guessing it might sound muffled if you played back the recordings of your voice as picked up by your palms resting on a keyboard. But, it might be good enough data that we can use some signal processing algorithm to tune our noisey, but high-quality, data for the user's voice.)