• This repository has been archived on 14/Oct/2022
  • Stars
    star
    143
  • Rank 257,007 (Top 6 %)
  • Language
    Python
  • License
    Apache License 2.0
  • Created about 4 years ago
  • Updated about 2 years ago

Reviews

There are no reviews yet. Be the first to send feedback to the community and the maintainers!

Repository Details

Moved to https://github.com/k2-fsa/icefall

snowfall

DEPRECATED: See Icefall instead.

This repo is deprecated in favor of its successor project Icefall.

About

Snowfall is an early draft of what will eventually be "icefall", the official recipes associated with k2 and lhotse. At the moment it is some early drafts of recipes, that we'll use for debugging and collaboration while the overall shape of the project becomes clearer.

Diagnostics

Our diagnotics are automatically collected using TensorBoard. You can inspect it locally by running:

$ tensorboard --logdir <exp_dir>

And then entering the url localhost:6006 in your browser (it is possible to change the port with --port option). When running the expts on a remote server, use port forwarding with SSH (ssh -L 6006:localhost:6006 user@address) so that your browser can connect to tensorboard.

Some noteworthy tensorboard options:

  • "Toggle all runs" to enable/disable all of the layers' plots
  • "Ignore outliers in chart scaling" on/off is useful depending on the plot;
  • "Tooltip sorting method" descending/ascending/closest, depending on the plot;
  • I usually set smoothing to 0;
  • I added a plot for "epoch" so it's easy to check at which step an epoch starts/ends.

Finally, Google added a service called "tensorboard.dev" where you can host your tensorboard data to show the exp to others. To do that, use the command: tensorboard dev upload --logdir <exp_dir> and follow the instructions in the terminal.

More Repositories

1

sherpa-onnx

Speech-to-text, text-to-speech, speaker recognition, and VAD using next-gen Kaldi with onnxruntime without Internet connection. Support embedded systems, Android, iOS, Raspberry Pi, RISC-V, x86_64 servers, websocket server/client, C/C++, Python, Kotlin, C#, Go, NodeJS, Java, Swift, Dart, JavaScript, Flutter, Object Pascal, Lazarus, Rust
C++
3,307
star
2

k2

FSA/FST algorithms, differentiable, with PyTorch compatibility.
Cuda
1,117
star
3

sherpa-ncnn

Real-time speech recognition and voice activity detection (VAD) using next-gen Kaldi with ncnn without Internet connection. Support iOS, Android, Linux, macOS, Windows, Raspberry Pi, VisionFive2, LicheePi4A etc.
C++
1,001
star
4

icefall

Python
912
star
5

sherpa

Speech-to-text server framework with next-gen Kaldi
C++
539
star
6

libriheavy

Libriheavy: a 50,000 hours ASR corpus with punctuation casing and context
Python
166
star
7

fast_rnnt

A torch implementation of a recursion which turns out to be useful for RNN-T.
Python
136
star
8

text_search

Some fast-ish algorithms for batch text search in moderate-sized collections, intended for data cleanup
Python
53
star
9

kaldifst

Python wrapper for OpenFST and its extensions from Kaldi. Also support reading/writing ark/scp files
C++
41
star
10

multi_quantization

Python
40
star
11

next-gen-kaldi-wechat

32
star
12

kaldi-decoder

Decoders from Kaldi using OpenFst
C++
18
star
13

colab

Colab notebooks for Next-gen Kaldi
Jupyter Notebook
14
star
14

analyze_diagnostics

Scripts for analyzing the output of icefall's "diagnostics.py" code (--print-diagnostics=True option)
Perl
5
star
15

k2-fsa-www

Source for next-gen Kaldi home page.
JavaScript
4
star
16

sherpa-torch-cpp-makefile-example

C++
4
star
17

divide_lm

Python
4
star
18

sherpa-onnx-go

sherpa-onnx Go package for speech recognition without network access, supporting Linux, macOS, Windows
Go
2
star
19

sherpa-onnx-go-windows

sherpa-onnx Go package for Windows
C
1
star
20

sherpa-ncnn-go-linux

sherpa-ncnn Go package for Linux
C
1
star