[
https://issues.apache.org/jira/browse/TIKA-94?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16507725#comment-16507725
]
Thejan Wijesinghe commented on TIKA-94:
---------------------------------------
WOW! this issue is so old. [~edwinyeozl] I'm very happy you are interested in
this feature. [~chrismattmann], A typical LSTM/GRU based speech recognition
model has around 120 million parameters, and training a model this large is
extremely computationally expensive, i.e. you need lots of GPUs. I already have
a working setup that does real time speech recognition, the one Mozilla has
trained and it has achieved a WER of 6.5% on LibriSpeech’s test-clean set, this
is close to human level ac curacies(human level WER is 5.83%), this is a
tensorflow setup, I can integrate this with Tika as an external docker setup as
an initial effort of including speech recognition in Tika, but supporting this
with DL4J could take some time. WDYT?
> Speech recognition
> ------------------
>
> Key: TIKA-94
> URL: https://issues.apache.org/jira/browse/TIKA-94
> Project: Tika
> Issue Type: New Feature
> Components: parser
> Reporter: Jukka Zitting
> Priority: Minor
> Labels: new-parser
>
> Like OCR for image files (TIKA-93), we could try using speech recognition to
> extract text content (where available) from audio (and video!) files.
> The CMU Sphinx engine (http://cmusphinx.sourceforge.net/) looks promising and
> comes with a friendly license.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
