Minimal set of scripts for training language and acoustic models for the speech recognition task. Training pipeline includes the following stages:

1. Character-based RNN language model

2. CNN-RNN acoustic model with CTC loss

3. Character-based RNN language model and CNN-RNN acoustic model with RNN-T loss

4. Fine-tuning with Reinforcement Learning and RNN-T loss

The following table shows the results for Russian Open Speech To Text (STT/ASR) Dataset.

• PyTorch >= 1.3 (with bug fix #27460)
• torch-edit-distance
• warp-rnnt

Acoustic models based on the log mel filterbanks with 40 filters of size 25ms, strided by 10ms.

• features.py - extract features of utterances listed in manifest file

Language model is character-based and not case sensitive.

• utterances.py - extract transcriptions of precomputed utterances

Pre-processed datasets:

• ru_open_stt_wav

Pre-trained models:

• ru_open_stt_models

There are outdated kernels with small training subsets:

• open-stt-lm

• open-stt-ctc

• open-stt-rnnt

• open-stt-rl