Word error rateby training round · illustrative
Scored by condition, not on average
Numbers come from the paper.
- Speech
- [N] h
- Speakers
- [N]
- Test set
- Held-out
- Metric
- WER
Method
From question to working system
- 01
Define
Where Pidgin speech AI fails today
- 02
Collect
Consented speech from real speakers
- 03
Train
Mini and Large, same data
- 04
Field-test
Calls, markets, noisy rooms
- 05
Measure
WER by region, noise, code-switching
- 06
Refine
Fix, retrain, publish what changed
Data
Consented, paid, removable.
Test speakers never appear in training. [confirm]
@misc{tok234,
title = {Tok234: speech recognition for Nigerian Pidgin},
author = {[Authors]},
year = {[Year]},
publisher = {Tala Labs},
url = {https://talaresearch.com/tok234}
}Show the work.
Read it, reproduce it, tell us what we missed.