MisaOgura

Miho’s tech test

First of all...

Well done on completing the bootcamp! I'm proud of you & glad you had a great time.

The project looks pretty good for a first tech test ;) You should be proud of yourself too!

Hopefully below feedback is useful... happy to jump on a call to discuss if you want too :)

MisaOgura

mkdir -p pkg
rm -rf pkg/kaldi_5.5
mkdir -p pkg/kaldi_5.5
cd pkg/kaldi_5.5                                                    \
	  && wget https://github.com/kaldi-asr/kaldi/archive/36f6dbf9a465c7ec94626a3f21debfb2c3483e80.tar.gz                                                 \
	  && tar --strip-components=2 -xvzf *.tar.gz kaldi-36f6dbf9a465c7ec94626a3f21debfb2c3483e80/src     \
	  && tar --strip-components=1 -xvzf *.tar.gz kaldi-36f6dbf9a465c7ec94626a3f21debfb2c3483e80/COPYING \
	  && touch kaldi.mk && make clean distclean && rm kaldi.mk                 \
	  && rm base/version.h                                                     \
	  && rm -rf *.tar.gz

MisaOgura

_________________________________________________________________ test_fetch_audio_stream _________________________________________________________________

self = <urllib.request.HTTPSHandler object at 0x7efee7794dd8>, http_class = <class 'http.client.HTTPSConnection'>
req = <urllib.request.Request object at 0x7efee7794e48>, http_conn_args = {'check_hostname': None, 'context': None}, host = 'youtube.com'
h = <http.client.HTTPSConnection object at 0x7efee7794f60>

    def do_open(self, http_class, req, **http_conn_args):
        """Return an HTTPResponse object for the request, using http_class.

MisaOgura

Milestone 1: How does the amount, variety and quality of training data effect accuracy of the model?

We can retrain our existing models with more / different data to assess what are the possible improvements that can be made. The preparation of the data for training and findings we make should have a longer application than just the current Kaldi based system (i.e. if we ever move to a seq2seq system).

Evaluation Datasets

We have now got multiple evaluation datasets to benchmark our system, we also

MisaOgura

SAD (Speech Activity Detection) - Viterbi decoding in Kaldi following VAD

Looking at detect_speech_activity.sh, these are the arguments/files/options that need to be provided to decode_sad.sh.

Arguments

1. graph_dir
   • graph_dir/HCLG.fst: An FST calculated with prepare_sad_graph.py
   • options: --min-silence-duration=0.03 --min-speech-duration=0.3 --max-speech-duration=10.0 --frame-shift=0.01

MisaOgura

Things we can try with GMM (suggested during a conversation with Jana)

• Increase the training iteration to few hundreds

  • Possible stopping criteria: min variance (keep an eye out for collapsing variance)

• Increase the number of Gaussian curves within each model

  • Currently init 32, diag 30 and full 20
  • Try init 258, diag 128 and full 64?

• Increase the size of MFCC

MisaOgura

# Choose a base image
FROM jupyter/scipy-notebook

# Information about the mainainter
LABEL maintainer="Misa Ogura <misa.ogura01@gmail.com>"

# Install TensorFlow & Keras
RUN conda install --quiet --yes \
    'tensorflow=1.3*' \
    'keras=2.0*' && \

MisaOgura

// scope chain: { name: undefined }
var name = 'John'
// scope chain: { name: 'John' }

function greet (name) {
  // [[scope]]: { name: 'John' }
  
  // scope chain: { name: 'John', outerScope: { name: 'John' } }
  return (function () {
    // [[scope]]: { name: 'John', outerScope: { name: 'John' } }

MisaOgura

var a = 'global'

function outer () {
  var b = 'outer'
  
  return function inner () {
    var c = 'inner'
    
    console.log('a:', a)
    console.log('b:', b)

MisaOgura

var a = 'global'

function outer () {
  var b = 'outer'
  
  return function inner () {
    var c = 'inner'
    
    console.log('a:', a)
    console.log('b:', b)