Josh Meyer

Feature request: hotword boosting for phrases

1

currently you can't boost phrases, only words split on whitespace. We should be able to boost phrases, too

Feature request: include raspberry pi install instructions in docs

3

STT.readthedocs.io is missing instructions on how to install from release page on raspberry pi's

Feature request: make scorer training a single script

1

Currently, to train a scorer you need to perform two key steps, after you have a cleaned text corpus: 1. train a KenLM model with `STT/data/lm/generate_lm.py` 2. package the model...

If you follow the steps in https://stt.readthedocs.io/en/latest/LANGUAGE_MODEL.html, you will not be able to train an LM because there is a dependency to *compiled* KenLM binaries. This is not at all...

it may be the case you run multiple training runs with the same checkpoints. each time you start a new training run, the `flags.txt` file gets overwritten. It would be...

We can point to the requirements for the base NVIDIA docker image, and give people an idea of what *should* work and what *definintely won't* work: From NVIDIA's TF container...

**Is your feature request related to a problem? Please describe.** Yes. I find myself saving training logs to a text file (what I consider) a hacky style. On a server,...

Feature request: Allow for easy transcript processing at training time

3

Scenario: I have a large dataset where all transcripts are in ALL CAPS, but the alphabet I want to use (i.e. fine-tune `v1.2.0`) is in lower case. Current solution: I...

Feature request: Add docs for using Webdataset in training

1

the use of `webdataset` in training was added in release `v1.2.0`, but we don't have docs for it's usage yet in `stt.readthedocs.io`

TTS has some nice gifs in the docs [here](https://tts.readthedocs.io/en/latest/inference.html), and STT could easily have some too. A gif for stt-model-manager and a gif for using the python client would be...