UA - kaldi instructional

Submission: Week 10

mybb@mybb.ru (CathyWalsh) — Fri, 30 Mar 2018 22:24:00 +0300

https://drive.google.com/file/d/1xaK3EI … sp=sharing

Submission: Week 8

mybb@mybb.ru (zupon) — Fri, 16 Mar 2018 21:57:49 +0300

Masha's submission:

https://drive.google.com/file/d/1oqtpLz … Nbjm3/view

Week 9 Agenda

mybb@mybb.ru (Michael Capizzi) — Fri, 09 Mar 2018 19:17:38 +0300

0. Questions/Concerns
1. Look ahead to final week and 10.HW: create your own experiment
2. Decoding at 50k feet
3. Comparing results
4. Lattices
5. decoding hyperparameters
A. `lmwt`
B. `beam`
C. `max_active`
6. using `analyze_*.log`s

Week 8 Agenda

mybb@mybb.ru (Michael Capizzi) — Fri, 09 Mar 2018 01:10:12 +0300

0. Questions/Concerns
1. update codebase for weeks 9 and 10
A. `sudo git pull`
- if conflicts:
1.`sudo git add` all files listed in output
2. `sudo git commit -m "before pull"`
3. `sudo git pull`
2. The **full** ASR pipeline
3. Look ahead to final week and 10.HW: create your own experiment
4. Review Case Studies
5. HCLG composition
6. Next week
A. Our first WER/SER results!

Week 7 Agenda

mybb@mybb.ru (Michael Capizzi) — Thu, 01 Mar 2018 18:57:44 +0300

0. Questions/Concerns
1. Remaining Schedule
2. Pull updates
A. `cd /scratch/kaldi`
B. `sudo git pull`
3. FSTs at 50k feet
4. setup HW template
5. look at `fst_manipulate.py`
6. FSTs in `kaldi`
A. `HCLG` --> after break
B. `G` --> today

Week 6 Agenda

mybb@mybb.ru (Michael Capizzi) — Fri, 23 Feb 2018 19:04:00 +0300

0. Questions/Concerns
1. Conserving AUs
2. Acoustic Modeling at 50k feet
A. train a decision tree to "state-tie" certain triphones together
B. train an HMM for each triphone where the transition probabilities are modeled by GMM (`pdf`)
- typically 3-5 states
- way of dealing with dipthongs?
C. given a frame, find the most representative `pdf`
3. Three Acoustic Models
A. Monophones
- quality of model not important
- poor alignment, but a "good start" that will be bootrapped
B. Triphones
- model context of triphones
- require "state tying" because the number of triphones is too large
- done by decision tree, asking "questions" about left and right context
C. Triphones + LDA_MLLT
- a transformation of the triphone model
- for what purpose?
4. Highlights of `exp` dir
A. the `mdl` file
5. Intuitions about acoustic modeling
A. number of Gaussians
B. number of leaves
5. Next week
A. openFST: http://www.openfst.org/twiki/bin/view/FST/WebHome
B. revisit our `ARPA` language model as an `FST`

Submission: Week 5

mybb@mybb.ru (zupon) — Fri, 16 Feb 2018 09:27:36 +0300

https://drive.google.com/file/d/1aeiaIW … sp=sharing

Week 5 Agenda

mybb@mybb.ru (Michael Capizzi) — Thu, 15 Feb 2018 20:15:23 +0300

0. Questions/Concerns
1. Case Study Reviews
2. Acoustic Modeling at 50k feet
A. See https://www.eleanorchodroff.com/tutoria … ncept.html
3. Three hierarchical acoustic models
A. Monophones
B. Triphones
1. decision tree
2. HMM
C. Triphones with LDA
3. Next Week

Week 4 Agenda

mybb@mybb.ru (Michael Capizzi) — Tue, 06 Feb 2018 19:41:23 +0300

1. Questions/Concerns
A. What to do when stuck in "networking"
B. What happens if you mess up entering your password
2. Quick review of `kaldi_config.json`(leftover from last week)
A. easiest way to edit?
3. Quick review of `data` dir (leftover from last week)
A. See 3.2
4. `MFCC` 50k foot view
A. See http://haythamfayek.com/2016/04/21/spee … rning.html
B. Best way to visualize?
5. `MFCC` config file
6. review of `mfcc` dir
A. `ark` v `scp`
7. Set up HW template
A. `python` kernel, can use `shell` cell with `%%bash`
B. Review of available methods
8. examine `mfcc` vectors
9. Next week
A. 4.HW
B. Week 5 ==> building acoustic models

Submission: Week 3

mybb@mybb.ru (damianiji) — Fri, 02 Feb 2018 22:05:35 +0300

https://drive.google.com/file/d/1lyEIAr … sp=sharing

Week 3 Agenda

mybb@mybb.ru (Michael Capizzi) — Tue, 30 Jan 2018 19:27:19 +0300

1. Questions/Concerns
2. AUs
a. if you don't see ~2300 AUs, click NEED MORE from your Dashboard; request 2304 AUs and explain you are part of the `kaldi` course
3. Pruning
a. What is the "meaning" of the threshold value?
- minimum unigram probability?
- minimum difference between probability and backoff?
4. Review Week 2 HW
5. Takeaways from Week 2 HW
6. Confirming the correct config file
7. Checking for error messages
8. Review data dir
a. Don't forget `lang_test_tg`!
9. What to expect next week
a. MFCCs: Sections 10.1 - 10.5 in Holmes and Holmes

Submission Process: Updated

mybb@mybb.ru (Michael Capizzi) — Fri, 26 Jan 2018 23:46:36 +0300

1. File -> Download As -> HTML
2. Save HTML into your Arizona Google Drive
3. Make public (or the option that allows for anyone with a `arizona.edu`address). `view` rights are plenty. It still enables downloading.
3. Post the link in the Sticky Forum for that week's submission

Week 2 Agenda

mybb@mybb.ru (Michael Capizzi) — Thu, 25 Jan 2018 20:24:57 +0300

0. Backup to volume
a. Volume can be found in `/vol_c`
b. directories to back up
- `raw_data`, `data`, `mfcc`, `exp`
c. command to use
- `cp -r [dir_to_copy] [path_to_volume]`
1. General Questions/Comments
2. Big picture contribution of LM
3. Brief summary of language modeling process: from corpus to LM
4. Smoothing
a. Why we need it
b. Which approach to use
5. IRSTLM
a. manual: http://hermes.fbk.eu/people/bertoldi/te … manual.pdf
b. already compiled and code in `/scratch/kaldi/tools/irstlm/bin` (must view from *inside* `docker`!)
c. how to use (see notebook 2.1)
6. Next week's HW
a. submission
- `File -> Download As -> HTML`
b. copy template
7. Generating the probability of a sequence (see notebook 2.2)
a. Default situation (len(sequence) <= size(n-gram) and n-gram in LM)
b. Two special situations
- n-gram not in LM
- sequence is larger than n-gram
8. "ate the lion" v. "ate the mouse" problem
a. Why did it happen?
9. Impact of LM n-gram size
10. Impact of LM size
a. space
b. speed
c. alternative options
- pruning (IRSTLM manual, section 5)
- hyperparameter = ???
- rescoring
11. Intuitions about ARPA-style LMs
12. n-gram LM v. RNN
13. What to expect next week
a. Week 3 items
1. `kaldi_config.json` usage
2. Building the `data` directory
b. Week 2 HW
1. identifying a case study
2. reviewing others' case studies

Week 1 Agenda

mybb@mybb.ru (Michael Capizzi) — Wed, 17 Jan 2018 00:48:17 +0300

0. Skim paper describing dataset building process: https://www.danielpovey.com/files/2015_ … speech.pdf
1. Table 1
2. 2.2-2.3 Alignment
3. 2.4 Data Segmentation
4. Table 2
1. General Questions/Concerns
2. Librispeech Dataset
a. Librivox v. Project Gutenberg
- https://librivox.org/
b. See paper here describing the build process of dataset: https://www.danielpovey.com/files/2015_ … speech.pdf
- See 2.4 Data Segmentation
- See Table 1
- See Table 2
- See 2.2-2.3 Alignment
2. Review of "Data Files"
a. splits
- train: `dev-clean`
- test: `test-clean`
- language model: `3-gram.pruned.3e-7.arpa`
b. audio
- 8 kHz v 16 kHz v flac
- ~20% reduction in performance with 8kHz v 16kHz (https://www.superlectures.com/odyssey20 … on-systems)
- listen to some samples
- male v. female
- listen to some samples
c. segmented v. unsegmented
- split on "pause"
- pause = silence for more than X seconds
- silence = no signal > Y dB
d. phones
- see http://www.speech.cs.cmu.edu/cgi-bin/cmudict
- silence phones
d. lexicon
- see http://www.speech.cs.cmu.edu/cgi-bin/cmudict
- find stressed v. unstressed examples
3. out-of-vocabulary (OOV)
- see http://www.speech.cs.cmu.edu/tools/lextool.html
- see https://github.com/sequitur-g2p/sequitur-g2p
- (`tmux session=sequitur` on desktop for demo)
4. What to expect next week
- resources, see schedule: https://docs.google.com/document/d/1pXt … mWpKbIeyqc
- 2.1
- in shell using IRSTLM (manual in resource_files)
- using a toy corpus
- "real" language model will be built in Week 3
- 2.2
- in python
- "exploratory" with "case study"
- 2.HW
- due before Week 3 class

Week 0 Agenda

mybb@mybb.ru (Michael Capizzi) — Wed, 17 Jan 2018 00:34:58 +0300

hammond wrote:

I just ran notebook "0_1" and did run_sh with -n 4. It was a lot faster, something like 30-45 minutes.

Yes! Since we have 4 CPUs on our instance, you can do that. Anything larger than 4 probably won't make a difference though.

General Info

mybb@mybb.ru (Michael Capizzi) — Tue, 28 Nov 2017 02:18:32 +0300

Course Length
First meeting will be 1/12
Final meeting will be 3/23

Time and Location
We will meet from 12:00 to 1:45 on Fridays in Douglas 216 (the "conference" room).

Schedule
The Google Doc of the proposed schedule is here.

Repository
The repository we will be using is a branch (kaldi_instructional) off of the main kaldi github repository. It can be found here.

Atmosphere Setup and Usage
See this page for instructions on setting up Atmosphere instance.
**Note**: The screenshots likely won't render when viewing on Github....sorry.

Atmosphere Allocation
We should have been given 2304 allocation units (AU). If you do *not* see that much allocation on your dashboard, let Michael Capizzi know.
That equates to a 4 CPU machine running 24 hours for 24 days. So if you use the VMs and shut them off when not in use that'll give you compute for the whole month except for 6 days.

Docker Image
The docker image we will be using can be found here. If you are planning on using your own machine, you'll need to access this image. See `Pulling from Dockerhub` in the project `README`.

General Usage
See the project README for instructions on using the `docker` `container` and `jupyter notebook`.