• Predictive models of modified speech: Three new objective measures designed to predict the subjective intelligibility in noise of speech (natural recorded or synthetic) that has been algorithmically modified.
  • Context estimator for speech output: A speech pre-enhancement method based on automatically recognising the speech, and matching the recognised text to that of the original message.
  • Live and recorded speech modifier: Seven techniques for modifying the speech signal to make it more intelligible in noise by exploiting audio and signal properties, and/or adoping modifications made by human talkers in intelligible speech styles (e.g. clear speech, Lombard speech), either independently of, or taking into account the time-frequency properties of, the noise masker.
  • MTRANS: A tool for annotating multi-channel speech. The software provides visual and aural flexibility in presentation of information to aid in the transcription of multi-party conversations.



(Please see the readme files for details of how to obtain the corpora)
  • DiapixFL: Six pairs of bilingual English-Spanish (L1-L2) and six pairs of bilingual Spanish-English talkers recorded while performing a “spot-the-difference” task.
  • Hurricane: Single male native British English talker recorded producing three speech sets (Harvard sentences, Modified Rhyme Test, news sentences) in quiet and while the talker was listening to speech-shaped noise presented at 84 dB(A).
  • Acted Clear Speech: Single male native British English talker recorded producing 25 TIMIT sentences in 5 conditions, two natural: (i) quiet, (ii) while the talker listened to high-intensity speech-shaped noise, and three acted: (i) as if to a non-native listener, (ii) as if to a computer speech-recognition system, (iii) as if to an infant. Accompanied by hand-corrected automatic phone-level transcription.
  • Sharvard: Two native Spanish talkers (one male, one female) recorded producing 720 Spanish sentences designed to be the Spanish equivalent of the English language Harvard sentences (thus phonetically balanced across sets of ten sentences).