ICSI Broadcast News Project

---

August Trainings Progress Page

There are several critical trainings in progress this month, August, 1998. This page will be kept up to date with the latest results.

---

Align 3 'Kitchen sink' (train24)

The latest set of alignments, align3, was installed at ICSI on Aug 21. These alignments are based on the output of a combined RNN/ModSpecGram acoustic model, so should be much better suited to ModSpec training. We immediately set to training a 4000 HU net for comparison:

Epoch	Date	Align 3 - msg1N-4k			Comparison (align2, train18)			Comments
		LrnRt	CV FA%	WER%	LrnRt	CV FA%	WER%
1	aug22 09:21	0.008	60.60		0.008	59.33	49.9
2	aug23 03:11	0.008	62.13		0.008	60.89	47.7
3	aug23 20:56	0.008	63.28		0.008	61.62	47.0
4	aug24 14:42	0.008	63.45		0.008	62.02	45.5
5	aug25 08:38	0.004	66.23	40.7	0.004	64.83	41.3	BN98I models
6	aug26 02:38	0.002	67.93		0.002	66.43	38.5
7	aug26 20:43	0.001	69.03		0.001	67.41	37.1
8	aug27 14:43	0.0005	69.72		0.0005	68.12	36.4
9	aug28 08:40	0.00025	70.08	35.6	0.00025	68.52	35.3	BN98I models
10					(training complete)
		combo w/RNN1		29.5	combo w/RNN1		29.3	BN98I models
		.. + 27hyp		27.6	.. + 27hyp		27.2

---

'Kitchen sink' training (train18)

This is the large (4000HU) net we are training to give to Cambridge to use in their next round of alignments. It is based on the modulation spectrogram (msg1) features, with a 28x9 input window. This is similar to our best-performing msg net so far, except that it is trained to the new 'align2' labels.

Epoch	Date	'Kitchen Sink'			Comparison (4000HU msg1N)			Comments
		LrnRt	CV FA%	WERR%	LrnRt	CV FA%	WERR%
1	aug01 12:53	0.008	59.33	49.9	0.008	56.58
2	aug02 06:45	0.008	60.89	47.7	0.008	57.83
3	aug03 00:37	0.008	61.62	47.0	0.008	58.47
4	aug03 18:35	0.008	62.02	45.5	0.008	59.39
5	aug04 12:31	0.004	64.83	41.3	0.008	58.97
6	aug05 06:26	0.002	66.43	38.5	0.004	61.87
7	aug06 00:24	0.001	67.41	37.1	0.002	63.43
8	aug06 18:31	0.0005	68.12	36.4	0.001	64.62
9	aug07 12:31	0.00025	68.52	35.3	0.0005	65.34
10		(training complete)			0.00025	65.74	37.8
		combo w/RNN1		29.3	combo w/RNN1		29.9

---

BAT for plp (train20)

Since we're throwing everything we've got at the problem this week, we are also training up a large plp12N net on all the data. We already trained a (13x9):4000:54 plp12N net to the align1 labels on half the data, so we used the net from the 4th iteration of this training (bn/experiments/fosler/train5) as the starting point for this net, and fixed the learning rate to ramp down, as if starting with epoch 6 of the previous training (since epoch 5 gave a negligable CV improvement). Comparison is against the equivalent iterations from the align1-half training.

E	Date	plp12N-4kHU-all-a2			plp12N-4kHU-half-a1			Comments
		LrnRt	CV FA%	WERR%	LrnRt	CV FA%	WERR%
6	aug04 20:19	0.004	63.49	39.9	0.004	62.55
7	aug05 17:13	0.002	65.29	37.8	0.002	63.68
8	aug06 14:30	0.001	66.20	36.5	0.001	64.52	36.0
9	aug07 11:30	0.0005	66.88	35.1	0.0005	65.06
10	aug08 08:25	0.00025	67.47	34.3	0.00025	65.34	35.0
11	aug09 05:14	0.000125	67.87	33.7	0.000125	65.48
12		0.000063			0.000063	65.55	34.4
		combo w/RNN1		29.7	combo w/RNN1		30.1

---

Align2 comparison training (train17)

We have only recently started working with the second set of alignments, generated by Cambridge using improved pronunciation modelling and a combination of more acoustic models. This training is a duplication of our 'standard' msg1N 2000HU training, but based on the align2 alignments. The corresponding figures for the align1 training are shown as comparison.

Epoch	Date	align2			align1			Comments
		LrnRt	CV FA%	WERR%	LrnRt	CV FA%	WERR%
1	jul31	0.008	57.35	53.3	0.008	56.39	51.9
2		0.008	58.32	51.3	0.008	57.60	50.5
3		0.008	59.17	48.7	0.008	58.05	48.5
4	aug01	0.008	59.86	48.6	0.004	61.05	44.5
5		0.008	59.88	48.2	0.002	62.59	42.2
6	aug02	0.004	62.87	42.8	0.001	63.73	39.9
7		0.002	64.26	40.3	0.0005	64.39	39.2
8		0.001	65.32	39.3	0.00025	64.75	39.4
9	aug03	0.0005	65.99	38.6	0.000125	64.95	38.8
10		0.00025	66.29	38.6	0.000063	65.04	38.8
11		(training finished)			0.000031	65.08	38.6
12					0.000016	65.09	38.5
13					0.000008	65.10	38.5
		combo w/RNN1		30.4	combo w/RNN1		30.1
		+RNN, 27hyp		28.5	+RNN, 27hyp		27.9

---

Cepstra-for-modspec training (train19)

Per-utterance normalization gave a big win on plp features, but had negligable effect on msg features. Other analysis reveals that, for the very shortest utterances, msg features did particularly well until they were normalized. One hypothesis is that because the msg features correspond almost directly to energy in given spectral bands, they might be strongly bimodal, corresponding to the two modes of voiced speech and silence (particularly in longer utterances that contain pauses). Hence, simple normalization could be very unfortunate, and depend critically on the proportion of silence in an utterance.

Using the cepstral transform on the modspec features has proven to make little or no difference in Brian's experiments with NUMBERS, but it might at least alleviate the bimodality of feature dimensions (if it exists). Since Brian had already written code to calculate cepstral features, it was easy to start a training on them (train19), whose progress is charted below.

This is a 2000HU net trained on 28-element feature vectors - full-order cepstral transforms of both msg1 spectra, followed by per-utterance normalization (msg1cepN). Training labels are from align2. Comparison net is the plain msg1N-align2-2kHU from above.

E	Date	msg1cepN			msg1N			Comments
		LrnRt	CV FA%	WERR%	LrnRt	CV FA%	WERR%
1	aug01 01:10	0.008	56.66		0.008	57.35	53.3
2	aug01 13:42	0.008	58.22		0.008	58.32	51.3
3	aug02 02:!5	0.008	58.82		0.008	59.17	48.7
4	aug02 14:47	0.008	59.65	47.7	0.008	59.86	48.6
5	aug03 03:21	0.008	59.79	47.3	0.008	59.88	48.2
6	aug03 16:00	0.004	62.71	42.4	0.004	62.87	42.8
7	aug04 05:06	0.002	64.10	39.9	0.002	64.26	40.3
8	aug04 17:58	0.001	65.02	38.7	0.001	65.32	39.3
9	aug05 06:31	0.0005	65.60	38.2	0.0005	65.99	38.6
10	(training halted)				0.00025	66.29	38.6
11					(training finished)
		combo w/RNN1		30.2	combo w/RNN1		30.4

Back to ICSI BN Home Page