Help Improve Open Source Speech Recognition

kmaclean · Post by **kmaclean** » March 2nd, 2007, 9:25 am

Hi,

I am the admin for the VoxForge project. VoxForge collects transcribed speech audio for the creation of Acoustic Models for use with Open Source Speech Recognition Engines ('SRE's). An Acoustic Model is basically a file that contains the statistical representations of sounds that make up the words in a large corpus of spoken audio. Currently, most Acoustic Models included with Open Source SREs are closed source (i.e. they don't provide the source audio because of license restrictions). VoxForge hopes to address this problem.

We are looking for submissions of audio books to help us with our goal. We are looking for audio books in uncompressed format (i.e. before you compress your audio book to mp3 for submission to Librivox), up to a 48kHz sampling rate at 16 bits per sample. Please consider submitting your ebook to VoxForge.

We've set up a page on the VoxForge site (called Uploads) that allows you to submit your audio books to VoxForge using an FTP client of your choice.

thanks in advance,

Ken

Starlite · Post by **Starlite** » March 2nd, 2007, 9:42 am

Great idea Ken but I'm afraid once a project is in the catalogue, most of us dump the raw, uncompressed files as they are sooo huge and we have limited space on our hard drives.

Will keep it in mind for the future though.

Esther

kayray · Post by **kayray** » March 2nd, 2007, 10:04 am

It's easy enough to take mp3s and re-convert them to .wav. Feel free to use any of our books, Ken!

Post by **hugh** » March 2nd, 2007, 10:13 am

hi ken contacted me about this... i thought we might be able to help... we could just put out a request here for say 20 people to offer up their next recordings in wav?

I'll contirbute one or 2...

kayray · Post by **kayray** » March 2nd, 2007, 10:18 am

Would single chapters be useful, or do you need entire complete books? (Sorry, haven't read your project page yet)

kristin · Post by **kristin** » March 2nd, 2007, 11:07 am

I still have the wav files from most of the recordings I've done (at least since I got my new computer.) Definitely, from two of my solo projects and my duet with Kara.

kmaclean · Post by **kmaclean** » March 2nd, 2007, 11:09 am

Starlite wrote:Great idea Ken but I'm afraid once a project is in the catalogue, most of us dump the raw, uncompressed files as they are sooo huge and we have limited space on our hard drives.

Will keep it in mind for the future though.

Esther

Hi Esther,

thanks for the reply,

Even a portion of what you recorded would be helpful. We need audio from as many different people, reading as much varied text as possible. We are looking for uncompressed audio because it provides the best quality audio for acoustic model creation.

Ken

kmaclean · Post by **kmaclean** » March 2nd, 2007, 11:22 am

kayray wrote:It's easy enough to take mp3s and re-convert them to .wav. Feel free to use any of our books, Ken!

Hi Kara,

MP3 compressed audio converted to wav is not the best audio for the creation of Acoustic Models for Speech Recognition. The compression introduces some noise that might affect the recognition process. Having said that, it may be 'good enough' for our purposes, and we are creating a sub-project within VoxForge (hopefully through Google Summer of Code) to explore this idea further, and are planning to use Librivox mp3 audio.

However, if we can get uncompressed audio, it would be better. Because Speech Recognition Engines work best to recognize the same type of speech audio (i.e. uncompressed) their Acoustic Models were trained with.

thanks,

Ken

kayray · Post by **kayray** » March 2nd, 2007, 11:25 am

I understand. And since random chapters are OK, it'll be easy to upload my chapters to you before I delete the .wavs.

kmaclean · Post by **kmaclean** » March 2nd, 2007, 11:32 am

kayray wrote:Would single chapters be useful, or do you need entire complete books? (Sorry, haven't read your project page yet)

Hi Kayray,

Any contribution would be greatly appreciated.

We need audio contributions from as many different people as possible (covering different dialects and regions), using different equipment (headset mics, desktop boom mics, etc. on laptops, desktops, using an audio card or usb pod) reading a variety of texts.

thanks,

Ken

kayray · Post by **kayray** » March 2nd, 2007, 11:57 am

Four chapters of "Persuasion" uploading now. Hope I followed all the protocol correctly :)

kmaclean · Post by **kmaclean** » March 2nd, 2007, 12:17 pm

kayray wrote:Four chapters of "Persuasion" uploading now. Hope I followed all the protocol correctly

Amazing! thank you so much.

When you get a chance, please include a README file with your submission. It's a bit tedious I know, but it will help with the classification of the your submitted audio. In addition, there are a few academics who interested in this project, and such information would help with their speech recognition research (especially microphone type).

Any feedback on improving the process would be greatly appreciated.

thanks,

Ken

kayray · Post by **kayray** » March 2nd, 2007, 12:24 pm

I did include a README... the upload won't finish for another hour, at least, so that file may not be on your server yet :)

The instructions were clear and straightforward. I remembered to amend my text files to include EVERY word that I spoke (librivox disclaimer, "end of chapter" etc). You didn't specify any filenaming conventions for the .wavs and textfiles, so I just went with something logical.

I am a passionate believer in open source projects, so I'm delighted to help with yours!

ductapeguy · Post by **ductapeguy** » March 2nd, 2007, 12:47 pm

What an awesome project. I've often felt that speech recognition is a neglected area of Open Source software. I will contribute what I can.

kmaclean · Post by **kmaclean** » March 2nd, 2007, 2:10 pm

kayray wrote:I did include a README... the upload won't finish for another hour, at least, so that file may not be on your server yet

The instructions were clear and straightforward. I remembered to amend my text files to include EVERY word that I spoke (librivox disclaimer, "end of chapter" etc). You didn't specify any filenaming conventions for the .wavs and textfiles, so I just went with something logical.

I am a passionate believer in open source projects, so I'm delighted to help with yours!

Sorry, I spoke too soon, I've got everything.

thanks for the submission!

Ken