A Comparison of Automatic Speech Recognition Systems

#1

Post by aberoham » Fri, Sep 14, 2018, 1:13 PM UTC

A Comparison of Automatic Speech Recognition Systems
blog.timbunce.org

Re: A Comparison of Automatic Speech Recognition Systems

#2

Post by singularity2001 » Fri, Sep 14, 2018, 2:16 PM UTC

So many unknown new players, almost all better than Nuance (thus Siri). Hard to believe though that those new companies are approaching Googles accuracy. As much as Google is worth being shunned for eroding privacy, their TTS always feels miles ahead of anything

Re: A Comparison of Automatic Speech Recognition Systems

#3

Post by bluGill » Fri, Sep 14, 2018, 2:53 PM UTC

I wonder how mozilla deepspeech compares https://github.com/mozilla/DeepSpeech

I'm sure it is not better than Google, but it seems to do fairly well in my limited tests.

Re: A Comparison of Automatic Speech Recognition Systems

#4

Post by gok » Fri, Sep 14, 2018, 2:55 PM UTC

This article seems to use "Google Text-to-Speech" in several places it means to say "Speech-to-Text"

Re: A Comparison of Automatic Speech Recognition Systems

#5

Post by gok » Fri, Sep 14, 2018, 2:59 PM UTC

bluGill wrote: post #3

I wonder how mozilla deepspeech compares https://github.com/mozilla/DeepSpeech I'm sure it is not better than Google, but it seems to do fairly well in my limited tests.

Note that Mozilla's "Project DeepSpeech" is just a training and recognition engine, and really an implementation of a specific design for LVCSR. A full recognition system requires statistical models, generally trained from very large sources of data, which you have to bring yourself.

Re: A Comparison of Automatic Speech Recognition Systems

#6

Post by raisedbyninjas » Fri, Sep 14, 2018, 3:15 PM UTC

Could this corpus be tested against Cortana? https://blogs.microsoft.com/ai/historic-achievement-microsof...

Re: A Comparison of Automatic Speech Recognition Systems

#7

Post by xerosanyam » Fri, Sep 14, 2018, 3:16 PM UTC

So happy to see Spext at #3 :)

Re: A Comparison of Automatic Speech Recognition Systems

#8

Post by gok » Fri, Sep 14, 2018, 3:44 PM UTC

raisedbyninjas wrote: post #6

Could this corpus be tested against Cortana? https://blogs.microsoft.com/ai/historic-achievement-microsof...

There is Microsoft's Azure speech-to-text API: https://azure.microsoft.com/en-us/services/cognitive-service...

But MSR's research system from that paper is probably quite different from what's deployed in Cortana/Azure, and Switchboard is a very different task from podcast transcription.

Re: A Comparison of Automatic Speech Recognition Systems

#9

Post by rorrr » Fri, Sep 14, 2018, 3:46 PM UTC

What ever happened to the seemingly amazing MS Research system that was demoed in 2012?

https://www.youtube.com/watch?v=Nu-nlQqFCKg

Re: A Comparison of Automatic Speech Recognition Systems

#10

Post by singularity2001 » Fri, Sep 14, 2018, 4:11 PM UTC

gok wrote: post #5

bluGill wrote: post #3

I wonder how mozilla deepspeech compares https://github.com/mozilla/DeepSpeech I'm sure it is not better than Google, but it seems to do fairly well in my limited tests.

Note that Mozilla's "Project DeepSpeech" is just a training and recognition engine, and really an implementation of a specific design for LVCSR. A full recognition system requires statistical models, generally trained from very large sources of data, which you have to bring yourself.

don't they provide (pre)trained models? someone on github does.