I wrote a blog detailing how you can transcribe all your voicemail messages and send the text via email with your voicemail message. The process will essentially convert the voice audio into text using PocketSphinx and send it in an email so you can read the voicemail without actually having to listen to the message!
The feature is similar to the voicemail transcription that Google Voice offers. I currently have this system working using Elastix 2.0.3 with FreePBX.
I have never had much luck with the Google version so my expectations were not so high. The installation guide is great but the bottom line is that the accuracy of this engine seems about the same as OCR was in 1989…if that:
Actual message left:
“Hello This is a test message. Let me see what you can do”
Transcribed Message:
“hello this tuesday to the fish all trying to to something that”
The actual transcription accuracy is definitely a difficult issue to tackle. There are ways you can “train” the acoustic model so it is more accurate with voicemail recordings. You can look at the PocketSphinx website for information on how to take some existing recordings and create your own acoustic model. Some acoustic models are not designed to recognize telephone audio, which can cause some discrepancies in the accuracy.
There are other acoustic models that you can try with this system that may give you more accurate results. VoxForge has Sphinx acoustic models that you can try out that might produce better results:
Although sometimes these transcriptions may not always be totally accurate, it can help give you an idea of what the call was about before actually listening to the message.
Let me start by listing what I’ve got. I know that this configuration is not ideal, but it works and I’m more familiar with Ubuntu.
Ubuntu 11.10
FreePBX 2.9.0.7
Asterisk 1.8.7.0
Apache 2.2.20
PHP 5.3.6
MySQL 5.1.58
Sendmail 8.14.4
Dovecot 2.0.13
ZendGaurd (the new Zend Optimizer Engine)
Now I get emails for voicemail notification, however there's no transcription. There doesn't appear to a transcription file. There appeared to be no errors during the component installations. And I'm not sure how to test Sphinxbase or Pocketsphinx. Any ideas on how i can fix this?
I used the same tutorial as listed above and followed it but I also don’t get the transcribed file created and therefore no transcription. I’m thinking the issue is with the transcript script and so trying to see where the issue might be. Will report any findings.