PodCastle is a service that enables users to find speech data that include a search term, read full texts of their recognition results, and easily correct recognition errors by simply selecting from a list of candidates.

As indices for spoken document retrieval, PodCastle provides the full text of speech recognition results for podcasts, individual audio or movie files on the web, and video clips on video sharing services such as Nico Nico Douga, YouTube, and Ustream. Users can read those full texts with a cursor moving in synchronization with the audio playback on a web browser. If a user finds a recognition error while listening, the user can easily correct the error by simply selecting from a list of candidates or typing the correct text on an efficient error correction interface. The resulting corrections can then be used not only to immediately be shared with other users and improve the spoken document retrieval performance for corrected speech data, but also to gradually improve the speech recognition performance by training our speech recognizer so that other speech data can be searched more reliably. This approach can be described as collaborative training for speech recognition.

Although we have used our own AIST's speech recognizer for Japanese language so far, we have started collaborating with the Centre for Speech Technology Research (CSTR), University of Edinburgh, and used the CSTR's speech recognizer for English language, which was developed under their EU projects "FP6 AMI" and "FP6 AMIDA".

• 2006/01 Started the project
• 2006/12 Released to the public
• The world's first speech retrieval project using crowdsourcing
• 2007/08 Interspeech 2007 papers
• Speech Recognition Research 2.0
• 2008/06 Press release (208/06/12)
• Reported in TV/web news, newspapers, etc.
• 2009/08 Supported video podcasts
• 2009/09 Interspeech 2009 paper
• 2011/10 Supported Nico Nico Douga, YouTube, and Ustream.tv
• 2011/10 Launched the English version
• 2011/10 Press release (2011/10/12)

The details of this research are available at http://staff.aist.go.jp/m.goto/PodCastle/.

Please send an e-mail to the following address.

• e-mail: podcastle-ml at aist.go.jp