Roadmap:
-
0.4.x: Fix some problems
- Absence of mdef_convert
- Make a general model conversion tool
- Allows us flexibility in model implementation for future
- Documentation deficiencies
- Lack of example for endpointer
- Lack of adapter for endpointer (WebAudio, etc)
- Lack of web audio format support
- Absence of mdef_convert
-
0.5.x: Improve performance/usability
- Phone input/output as IPA
- Optimize JSGF compiler
-
1.0: Finalize API
- ES6 module (possibly separate for Node vs. Web)
- Improve Endpointer/VAD
- Clearly define use cases and restructure API for them
- Live: feed data asynchronously, check results synchronously or emit events
- recording is real-time in a separate thread
- VAD/Endpointing should be done in the recording thread
- Feature extraction could be done there but this is not necessary
- decoding is not real-time, can be decomposed to microtasks
- recording is real-time in a separate thread
- Single/Batch: pass data with promise of result, emit progress events
- Good fit for async
- Live: feed data asynchronously, check results synchronously or emit events
- Easier support for observable/event type uses
- async/await/promise is actually not a great fit
- Support web audio formats directly
- Better solution for float vs. int in front-end
-
2.0: Improved modeling
- WFST search
- DNN acoustic models