nvidia-nemo

NVIDIA NeMo ASR

NVIDIA NeMo is a toolkit for conversational AI including several models for ASR.

Links:

• NeMo GitHub: https://github.com/NVIDIA/NeMo
• NeMo ASR Docs: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/intro.html
• NeMo ASR Models: https://catalog.ngc.nvidia.com/orgs/nvidia/collections/nemo_asr
• Language Modeling: https://docs.nvidia.com/deeplearning/nemo/user-guide/docs/en/main/asr/asr_language_modeling.html
• KenLM: https://github.com/kpu/kenlm

Tested with:

• Arm64 - Debian 11 - Python 3.9
• x86_64 - Debian 11 - Python 3.9

Train and use language models

The beam search decoders in NeMo support N-gram language models trained for example with the KenLM toolbox.
Use bash install_lm_tools.sh to get pre-built KenLM binaries and the required ctc-decoders. The wheels are not available for all Python version, so you might need to build them yourself (see: asr_language_modeling/ngram_lm/install_beamsearch_decoders.sh).

• Check-out the 'train' folder to learn how to build a simple n-gram LM in ~5 minutes
• Use python3 test_with_lm.py -h to see available options
• Read the Nvidia 'language modeling' docs (see above) to learn about beam width, alpha and beta parameters
• To run inference with LM use something like python3 test_with_lm.py -m "models/stt_en_conformer_ctc_small.nemo" -l "models/kenlm_custom.4gram" -w 4 -a 1.0 -b 2.0
• You can get a "naive" character error rate (CER) score for transcriptions by adding a txt-file for each wav-file with the target transcription (same name, just '.txt' ending) and using the --transcriptions [folder] argument