This PR integrates LODR (Level-Ordered Deterministic Rescoring) support from Icefall into both online and offline recognizers, enabling LODR for LM shallow fusion and LM rescore.
- Extended OnlineLMConfig and OfflineLMConfig to include lodr_fst, lodr_scale, and lodr_backoff_id.
- Implemented LodrFst and LodrStateCost classes and wired them into RNN LM scoring in both online and offline code paths.
- Updated Python bindings, CLI entry points, examples, and CI test scripts to accept and exercise the new LODR options.
* online-transducer: reset the encoder toghter with 2 previous output symbols (non-blank)
- added `reset_encoder` boolean member into the OnlineRecognizerConfig class
- by default the encoder is not reset
* pybind11, adding empty symbols for disabled modules (tts, diarization)
* reset_encoder, add default value (false) [pybind11]
* Use PROJECT_SOURCE_DIR rather than CMAKE_SOURCE_DIR to allow building as subdirectory
* Also use PROJECT_SOURCE_DIR instead of CMAKE_SOURCE_DIR in c/cxx api examples
* Only build examples by default when not building as subdirectory
* Do not suggest building binaries either
---------
Co-authored-by: user <user@mail.tld>
* adding ebranchformer encoder
* extend surfaced FeatureExtractorConfig
- so ebranchformer feature extraction can be configured from Python
- the GlobCmvn is not needed, as it is a module in the OnnxEncoder
* clean the code
* Integrating remarks from Fangjun
* Adding temperature scaling on Joiner logits:
- T hard-coded to 2.0
- so far best result NCE 0.122 (still not so high)
- the BPE scores were rescaled with 0.2 (but then also incorrect words
get high confidence, visually reasonable histograms are for 0.5 scale)
- BPE->WORD score merging done by min(.) function
(tried also prob-product, and also arithmetic, geometric, harmonic mean)
- without temperature scaling (i.e. scale 1.0), the best NCE was 0.032 (here product merging was best)
Results seem consistent with: https://arxiv.org/abs/2110.15222
Everything tuned on a very-small set of 100 sentences with 813 words and 10.2% WER, a Czech model.
I also experimented with blank posteriors mixed into the BPE confidences,
but no NCE improvement found, so not pushing that.
Temperature scling added also to the Greedy search confidences.
* making `temperature_scale` configurable from outside
