readme.md

Monolingual ASR model

Monolingual ASR model with full data

We trained a phoneme-based ASR model for each language of cv-lang10 with the same architecture that is based on a Conformer network consisting of 14 encoder blocks. The number of phonemes and training hours of the each language are in the following table.

Language	Language ID	# of phonemes	Train hours	Dev hours	Test hours
English	en	39	2227.3	27.2	27.0
Spanish	es	32	382.3	26.0	26.5
French	fr	33	823.4	25.0	25.4
Italian	it	30	271.5	24.7	26.0
Kirghiz	ky	32	32.7	2.1	2.2
Dutch	nl	39	70.2	13.8	13.9
Russian	ru	32	149.8	14.6	15.0
Swedish	sv-SE	33	29.8	5.5	6.2
Turkish	tr	41	61.5	10.1	11.4
Tatar	tt	31	20.8	3.0	5.7

• %PER

  Model	Model size	en	es	fr	it	ky	nl	ru	sv-SE	tr	tt	Avg.
  Mono. phoneme	90 MB	7.39	2.47	4.93	2.87	2.23	4.60	2.72	18.69	6.00	10.54	6.11

• %WER with 4-gram LM

  Model	Model size	en	es	fr	it	ky	nl	ru	sv-SE	tr	tt	Avg.
  Mono. phoneme	90 MB	10.59	7.91	15.58	9.26	1.03	8.84	1.62	8.37	8.46	9.75	8.14

Monolingual ASR model with low-resource data

For ablation study, the training data is divided into three scales to simulate different resource scenarios: 1 hour, 10 hours, and full data. Phoneme-based and subword-based models are both trained with this three scales of training data.

Language	Language ID	# of phonemes	# of subwords	Train hours	Dev hours	Test hours
Indonesian	id	35	500	20.8	3.7	4.1
Polish	pl	35	500	129.9	11.4	11.5

Phoneme-based

• %PER

  language	1 hour	10 hours	full data
  Indonesian	96.52	26.34	5.74
  Polish	86.01	30.38	2.82

• %WER with LM

  language	1 hour	10 hours	full data
  Indonesian	100	9.54	3.28
  Polish	99.98	13.86	4.97

Subword-based

• %WER without LM

  language	1 hour	10 hours	full data
  Indonesian	96.62	69.57	31.96
  Polish	98.41	90.98	19.38

• %WER with LM

  language	1 hour	10 hours	full data
  Indonesian	96.42	49.67	10.85
  Polish	98.38	59.43	7.12