Experiences with Google voice

Following a suggestion from Aarne, I found some Google service for speech input, but the experiments are not encouraging:

  1. I recorded Compute this into a mp4 file using QuickTime Player on the mac

  2. Converted it to flac using:

    sox compute.m4a compute.flac rate 16k

  3. And get into the service by:

    curl -H "Content-Type:audio/x-flac; rate=16000" "https://www.google.com/speech-api/v1/recognize?xjerr=1&client=chromium&lang=en-US" -F "myfile=@compute.flac

But got:

 `{"status":0,"id":"56bdb158dd66b25fc2e221364004e620-1","hypotheses":[{"utterance":"coffee lol","confidence":0.46219563}]}`

Other examples:

  • "I like pickles" ⇒ "I like turtles"

  • "The determinant of x" ⇒ "new york" (with confidence 0.88!)

  • "Compute this" ⇒ "coffee lol"

Of course I'm not a native English speaker, but I expected a better performance.

Comments

Script for experimenting with the Google speech recognizer

Submitted by Kaarel.Kaljurand on 5 January, 2012 - 15:16.

I wrote a script once that simplifies experimenting with the Google speech recognition service (tested on Linux):

I tried your examples and everything worked in quite a stable way except for "compute this".

In general, it would make sense to compile the GF grammar into a speech recognition grammar (e.g. JSGF) and pass it to the recognizer along with the audio. This would make sure that any recognition output can be generated/parsed by the original grammar. The Google server however does not support speech recognition grammars to my knowledge, so it would be better to use another speech recognition provider, e.g. one based on CMU Sphinx (http://cmusphinx.sourceforge.net/). Using the Google server from anything outside of Google Chrome or the Google Voice Search app on Android might be in violation with its terms of service anyway.

Kaarel

Hi Kaarel, I tried to

Submitted by jordi.saludes on 9 January, 2012 - 13:27.

Hi Kaarel,

I tried to reproduce your results but I don't know of a OSX equivalent for arecord.

I hope you would go to Goteborg this week: We should talk about it!

See you soon.

--Jordi Saludes

What links here

No backlinks found.