Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add the description of the diphone alignment system in PocketSphinx #30

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

Add the description of the diphone alignment system in PocketSphinx #30

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 31 additions & 0 deletions wiki/diphones.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,6 +5,35 @@ A *diphone* is the last part of one phoneme followed by the first part of
another. Either phoneme could be silence, and they can be the same phoneme.
Diphthongs include diphones in them.

There are 2 variants of the diphone alignment system in PocketSphinx.

The first one is **synthetic**, which builds diphone units automatically
from context-dependent phone units (triphones) after reading their
definition from `mdef` file. More precisely, it is done as follows:
for each possible pair of base phones, it tries to find 2 triphones:
one triphone that has the first base phone as main phone and the second
base phone as RC (right context), and another triphone that has the
first base phone as LC (left context) and the second base phone as
main phone. Then it takes the last 1 senone from the first found
triphone and the first 2 senones from the second found triphone.
Use `-diphones=synthetic` command line parameter to enable this variant.

The second one is **trained**, which uses pretrained acoustic model
where diphones are defined as context-independent units.
The model was trained on the "clean" subset of
[LibriSpeech](http://www.openslr.org/12/) ASR corpus and
contains 899 diphones. It requires the dictionary to use diphone
units as well. The version of CMU Sphinx `en-us` dictionary
with diphones was created with
[this](https://github.com/akreal/diphones/blob/master/scripts/dict.py)
script and it is also used automatically when the trained
diphones acoustic model is chosen. Use `-diphones=trained` command
line parameter to enable this variant.

Additionally, you can use `-diphones=yes` command line parameter,
which is currently an alias for the trained variant of
the diphones alignment system.

[This list of the top 4,800 words by frequency in English
speech](http://ucrel.lancs.ac.uk/bncfreq/lists/2_2_spokenvwritten.txt) was used
with [CMUDICT](http://www.speech.cs.cmu.edu/cgi-bin/cmudict) to create the
Expand All @@ -13,6 +42,7 @@ by approximate prevalence.

![diphones](/data/diphones.png)

```
UH_R 2.376%,
AH_N 2.083%,
T_SIL 1.863%,
Expand Down Expand Up @@ -1065,3 +1095,4 @@ ZH_V 0.003%,
ZH_W 0.003%,
ZH_Y 0.003%,
ZH_Z 0.003%.
```