With the need to cater to a global audience, there is a growing demand for applications to support speech identification/translation/transliteration from one language to another. This session aims at introducing the audience to the topic, learn the inner working of the AI/ML models and eventually how to quickly use some of the readily available APIs to identify, translate or even transliterate speech/text within their application.
4. Machine translation is the task of automatically converting source
text in one language to text in another language.
5. Top 10 languages*
81%
* English, Chinese, Japanese, French, German + Spanish, Portuguese, Russian, Italian, Korean
One Language: 95%
5%
Two or more languages
15. Detect
Bilingual dictionary
Transliteration
Multiple languages
See examples of human translated sentences using the input
word and find alternative translations for the input word
Llamando solo para comprobar.
I need you to check some files for me.
Necesito comprobar unos archivos.
Alternatives:
VERB
comprobar check, verify, test, prove, ascertain
revise check, review, inspect
verificar verify, check, verification
chequear check
NOUN
cheque check, paycheck, certificate
verificación verification, check, checking, verifying, credentials
comprobación check, checking, test, verification, physical, testing, verifying
18. bleue…
The
[0.1, -0.3,…,0.5]
La
[0.4, 0.7,…,0.3]
blue
[0.02, 0.4,…,0.91]
[0.2, 0.3,…,0.3]
maison
house
[0.1, 0.7,…,0.4] [0.1, 0.7,…,0.5]
is…
[-0.1, 0.4,…,0.8] [0.2, 0.3,…,0.1]
Attention algorithm
Final input matrix
[0.6, 0.02,…,0.7] [0.3, 0.2,…,0.01]
Model layer 1
Each word is modeled in context of the full sentence.
Model layer 2 to N
Multiple layers allow for better contextualization of
a given word as part of a whole sentence.
Attention layer
“Attention” layer (algorithm), defines word order translations
based on context.
Decoder
Final layer, the decoder, translates words with contextual
awareness for this particular language pair.
{
During training, the NN, creates a 500-
dimensions model of each word for a
given language pair:
• Word type (noun…)
• Singular/plural
• Gender
• Formality, ...
Note: examples only for illustration purposes.
Actual “dimensions” can be anything derived
by the NN after training
The NN, creates a 1000-dimensions
model of each word given the context
Translation layer of the NN has
learned word translations based on this
1000-dimensions sentence context
31. Let go of listening before
going to the nextto the next
listening
32. Français English
Let go of listening before
going to the nextto the next
listening
1. Upload
2. Train
3. Test
4. Deploy
1. Upload
33. Français English
Let go of listening before
going to the nextto the next
listening
2. Train
3. Test
4. Deploy
1. Upload1. Upload
2. Train
CUSTOM MODEL
34. Français English
Let go of listening before
going to the nextto the next
listening
1. Upload
2. Train
4. Deploy
3. Test CUSTOM MODEL
3. Test
2. Train
CUSTOM MODEL
35. Let go of listening before
going to the nextto the next
listening
1. Upload
2. Train
4. Deploy
3. Test CUSTOM MODEL
+4BLEU
SCORE
General Model: BLEU Score = 22
Custom Model: BLEU Score = 26
Français English
36. Let go of listening before
going to the nextto the next
listening
1. Upload
2. Train
3. Test
4. Deploy4. Deploy
37. CatID
Let go of the sheet before
going on the windon the wind
the sheet
Let go of listening before
going to the nextto the next
listening