[Speechsc] Continuous speech recognition in MRCP

[Tomáš Valenta <[e-mail]>]

Dear SpeechSC list members,

in our company we are implementing TTS and ASR solutions using MRCP. For 
ASR we would like to use the protocol not only for recognition of short 
utterances based on simple grammar; scheme

C->S: RECOGNIZE
S->C: IN-PROGRESS
S->C: START-OF-INPUT
S->C: RECOGNITION-COMPLETE (result)

but also for continuous speech recognition (e.g. minutes or tens of 
minutes, dictations, ...) with immediate results. Unfortunately there is 
no such approach in MRCPv2 specification draft. We thought about using 
following scheme:

C->S: RECOGNIZE (continuous)
S->C: IN-PROGRESS
S->C: START-OF-INPUT
S->C: IN-PROGRESS (partial_result_1)
...
S->C: IN-PROGRESS (partial_result_n)
C->S: STOP

Imagine an application for writing dictation so that user can see 
immediately what he said. The recognizer could be located on a remote 
machine.

The (continuous) parameter to the RECOGNIZE request could be a type of 
grammar (built-in language model specification, in fact) or a header 
value. Do you find this approach a clean solution? Or do not you think 
continuous speech recognition should be part of MRCPv2 specification?

Kindest regards and thanks for comments,
Tomas Valenta

PS. Originally we started discussing this topic with Arsen Chaloyan 
(author of UniMRCP) here:
https://groups.google.com/d/topic/unimrcp/pSaDbhHPh3M/discussion