ASR-aided transcription with segmented feedback training

1. An ASR-aided transcription system with segmented feedback training, the system comprising:

a transcription process manager configured to extract a first segment and a second segment from an audio input of speech uttered by a speaker; and

an ASR engine configured to:

operate in a first speech recognition mode to convert said first speech segment into a first text transcript using a speaker-independent acoustic model and a speaker-independent language model;

operate in a first training mode to create a speaker-specific acoustic model and a speaker-specific language model by:

adapting said speaker-independent acoustic model using said first segment and/or a proofread version of said first text transcript; and

adapting said speaker-independent language model using at least the proofread version of said first text transcript; and

operate in a second speech recognition mode to convert said second speech segment into a second text transcript using said speaker-specific acoustic model and said speaker-specific language model.

2. A system according to claim 1 wherein said ASR engine is configured to operate in a second training mode to:

adapt said speaker-specific acoustic model using said second speech audio segment and/or a proofread version of said second text transcript; and

adapt said speaker-specific language model using at least said proofread version of said second text transcript.

3. A system according to claim 2 wherein said ASR engine is configured to alternately operate in said second speech recognition mode and said second training mode for processing a plurality of segments of said audio input subsequent to processing said first segment.

4. A system according to claim 2 wherein said ASR engine is configured to cyclically process multiple speech segments from multiple speakers, where said ASR engine uses a different set of said speaker-specific models for each of said speakers.

5. A system according to claim 1 wherein said ASR engine is configured when in either of said speech recognition modes to determine a plurality of word time offsets for words in either of said segments, and when in said training mode to adapt said speaker-specific models using said word time offsets.

6. A system according to claim 5 and further comprising an Acoustic Model Training Dataset Preparator operative to

receive any of said segments, said proofread transcript, and said word time offsets,

divide said received segment and said proofread transcript into a plurality of audio and text pieces, and

create a training dataset of corresponding pairs of said audio and text pieces,

wherein said ASR engine is configured to use said training dataset when in any of said training modes to adapt any of said speaker-specific acoustic models.

7. An ASR-aided transcription method with segmented feedback training, the method comprising:

a) extracting a first segment and a second segment from an audio input of speech uttered by a speaker;

b) converting said first speech segment into a first text transcript using a speaker-independent acoustic model and a speaker-independent language model;

c) creating a speaker-specific acoustic model by adapting said speaker-independent acoustic model using said first segment and/or a proofread version of said first text transcript;

d) creating a speaker-specific language model by adapting said speaker-independent language model using at least said proofread version of said first text transcript; and

e) converting said second speech segment into a second text transcript using said speaker-specific acoustic model and said speaker-specific language model.

8. A method according to claim 7 and further comprising

f) adapting said speaker-specific acoustic model using said second speech audio segment and/or a proofread version of said second text transcript; and

g) adapting said speaker-specific language model using at least said proofread version of said second text transcript.

9. A method according to claim 8 and further comprising cyclically performing said e) converting, said f) adapting and said g) adapting for processing a plurality of segments of said audio input subsequent to processing said first segment.

10. A method according to claim 8 and further comprising alternatingly processing multiple speech segments from multiple speakers using separate sets of said speaker-specific models for each of said speakers.

11. A method according to claim 7 and further comprising:

determining a plurality of word time offsets for words in either of said segments; and

adapting said speaker-specific models using said word time offsets.

12. A method according to claim 11 and further comprising:

dividing any of said segments and said proofread transcript corresponding to said segment into a plurality of audio and text pieces;

creating a training dataset of corresponding pairs of said audio and text pieces; and

adapting any of said speaker-specific acoustic models using said training dataset.

13. A computer-readable recordable device encoded with a plurality of instructions that, when executed by at least one computer, perform a method comprising:

extracting a first segment and a second segment from an audio input of speech uttered by a speaker;

converting said first speech segment into a first text transcript using a speaker-independent acoustic model and a speaker-independent language model;

creating a speaker-specific acoustic model by adapting said speaker-independent acoustic model using said first segment and/or a proofread version of said first text transcript;

creating a speaker-specific language model by adapting said speaker-independent language model using at least said proofread version of said first text transcript; and

converting said second speech segment into a second text transcript using said speaker-specific acoustic model and said speaker-specific language model.

14. The computer readable recording device of claim 13 , wherein the method further comprises:

adapting said speaker-specific acoustic model using at least said second speech audio segment and/or a proofread version of said second text transcript; and

adapting said speaker-specific language model using at least said proofread version of said second text transcript.

15. The computer readable recording device of claim 13 , wherein the method comprises cyclically processing a plurality of segments of said audio input subsequent to processing said first segment.

16. The computer readable recording device of claim 13 , wherein the method further comprises:

determining a plurality of word time offsets for words in either of said segments; and

adapting said speaker-specific acoustic model using said word time offsets.

17. The computer readable recording device of claim 13 , wherein the method further comprises:

dividing any of said segments and said proofread transcript corresponding to said segment into a plurality of audio and text pieces;

creating a training dataset of corresponding pairs of said audio and text pieces; and

adapting any of said speaker-specific acoustic models using said training dataset.

18. The computer readable recording device of claim 13 , wherein the method comprises cyclically processing multiple speech segments from multiple speakers using separate sets of said speaker-specific models for each of said speakers.