#DeepSpeech in < 5 mins

open source sequence to sequence
speech recognition for everyone

 

Kathy Reid @KathyReid

PhD Candidate @3AInstitute, Australian National University & Open Source Voice Specialist @Mozilla

Canberra Python User Group, 4th March 2021

2021

7.85 billion
people

Source: https://www.worldometers.info/world-population/

7139
languages

Source: https://www.ethnologue.com/guides/how-many-languages

≈200

languages have speech to text support

Source: Internal Mozilla research

135 languages

55 of which are variants of
English, Spanish or Arabic

Source: Internal Mozilla research

so?

what about DeepSpeech?

inference or training

seq2seq

does not use phonemes

Vowels Monopthongs Dipthongs
iː
sheep		 ɪ
ship		 ʊ
good		 uː
shoot		 ɪə
here		 eɪ
wait
e
bed		 ə
teacher		 ɜː
bird		 ɔː
door		 ʊə
tourist		 ɔɪ
boy		 əʊ
show
æ
cat		 ʌ
up		 ɑː
far		 ɒ
on		 eə
hair		 aɪ
my		 aʊ
cow
Consonants p
pea		 b
boat		 d
tea		 t
dog		 tʃ
cheese		 dʒ
June		 k
car		 g
go
f
fly		 v
video
 θ
think		 ð
this		 s
see		 z
zoo		 ʃ
shall		 ʒ
television
m
man		 n
now		 ŋ
sing		 h
hat		 l
love		 r
red		 w
wet		 j
yes		 

alphabet.txt

contains the characters to train for

ⴰⵣⵓⵍ

ng’ombe

Cześć

Helô

≈201

languages have speech to text support

Thank you