WORKSHOP '96

The Results

This page collects all the results we have obtained during this workshop. Please refer to the Description of the experiments page for the complete description of the experiments.

The baseline system

We trained 4 MLP's, each with 2000 hidden units, rasta12(+delta+ddeltas) and delta+ddelta energy features (38 features total), 9 input frame window, and 56 outputs corresponding to the icsi phoneme set.
Each of the 4 MLP's were trained on a random set of sentences of SWB (cut0, cut1, cut2 and cut3). We tested the system on a random 300 sentences set from the sidev set. The probabilities are the averaged posterior probabilities of cut0+cut1+cut2+cut3 nets.

word entrance penalty = 15
lm scaling factor = 0.3 word entrance penalty = 4
lm scaling factor = 0.5

64.9 % 64.3 %

speaker stats speaker stats

Similar experiments have been carried out on a 4hr male training set. Recognition was performed on 240 utterances (only males).

word entrance penalty = 15
lm scaling factor = 0.6

63.6 %

speaker stats

Those experiments were performed using :

Dictionary : /export/n/ws96/SWB/Lex/rasta12-iter4.recog.noway.upn.pr0.5.lex
Phoneme file : /export/n/ws96/SWB/Lex/rasta12-iter4.recog.noway.phmodels
Priors : /export/n/ws96/SWB/Priors/rasta12-iter4.priors

The multiband system

The seven band experiment

Data sets being used are:

Training set: 8hr ICSI cut0
Cross-validation (cv) set: 10% of the training sentences
Test set: 300 random utterances from the dev-test

Parameters used are:

Features: PLP cepstra + delta_gain + delta_cepstra + delta_delta_gain + delta_delta_cepstra
Number of subbands: 7
Number of MLP hidden units: 2000
Number of MLP outputs: 56 (icsi phoneme set)

Software used:

QuickNet from ICSI
STRUT lattice-decoder from Mons.

Latest results:

Recognizer	Approximate frequency range(Hz)	Input layer size	Frame level error% (cv set)	word error rate% (test set)
Baseline (conventional)	0:4000	234	48.62	67.6
*MLP merging of*	...	...	...	...
all 7 subbands	0:4000	392	-	63.7
subbands 2-7	200:4000	336	-	64.1
subbands 1-6	0:2767	336	-	64.2

Recognizer	Approximate frequency range(Hz)***	Input layer size	Frame level error% (cv set)	word error rate% (test set)
subband 1	0-300	153	72.27	74.6
subband 2	200-670	207	66.53	72.6
subband 3	550-980	153	66.81	78.1
subband 4	820-1420	153	66.73	75.4
subband 5	1200-1984	153	65.11	74.0
subband 6	1670-2765	153	67.39	73.9
subband 7	2300-4000	153	70.21	74.5

*** Note each sub-band roughly covers two critical bands

Recent Multi-band results

Training set: 9019 sentences, male speakers, cut1
Test set: 240 male utterances
Floating point libraries used for training instead of fixed-point libraries used in the 7 sub-band experiments reported earlier.

Recognizer	Frequency range(Hz)	NN size	Frame level error %	word error rate %
subband 1	0-300	153:500:56	70.76	72.70
subband 2	200-670	207:500:56	61.23	68.30
subband 3	550-980	153:500:56	62.13	69.80
subband 4	820-1420	153:500:56	61.05	69.70
subband 5	1200-1984	153:500:56	57.80	67.80
subband 6	1670-2765	153:500:56	62.33	68.10
subband 7	2300-4000	153:500:56	66.56	71.00
.	.	.	.	.
Baseline	0-4000	234:500:56	42.65	60.90
MLP merging	0-4000	392:500:56	42.88	59.70
Linear merging	0-4000	.	.	65.10

7 sub-band model.

Comparison with the the earlier 7 band experiment:

Floating point libraries used in the training improved the word error rates 4-8 % (absolute values).
Best merging performance using non-linear MLP merging (59.7 % error)

The four band experiment

Data sets being used are:

Training set: 4hr ICSI male cut
Cross-validation (cv) set: 10% of the training sentences
Test set: 240 random utterances (only males) from the dev-test

Parameters used are:

Features: PLP cepstra with CMS + delta_cepstra + delta-log-energy + delta-delta-log-energy
Number of subbands: 4
Number of MLP hidden units: 500
Number of MLP outputs: 56 (icsi phoneme set)

Software used:

QuickNet from ICSI
STRUT lattice-decoder from Mons.

Recognition results

Recognizer	Approximate frequency range(Hz)	Input layer size	Frame level error% (cv set)	word error rate% (test set)
subband 1	0-900	234	59.7	68.8
subband 2	800-1660	234	57.9	68.5
subband 3	1500-2550	162	60.3	69.2
subband 4	2300-4000	162	67.2	69.8

The four bands were recombined at the state level using :

a untrained linear recombination strategy : sum of the log-likelihoods

word entrance penalty = 15
lm scaling factor = 0.6 61.4 %

a trained linear recombination strategy : ANN without hidden layer

word entrance penalty = 15
lm scaling factor = 0.6 61.0 %

We also have combined the four bands (recombined with the ANN) with the full band probabilities :

word entrance penalty = 15
lm scaling factor = 0.6

59.4 %

The speaking rate

word entrance penalty = 15
lm scaling factor = 0.6

63.5 %

The modulation spectral filters

word entrance penalty = 15
lm scaling factor = 0.6

62.4 %

The Chaf

The results obtained with the CHAF are reported here

Last modified on August 23, 1996

Christophe Ris <ris@cspjhu.ece.jhu.edu >

word entrance penalty = 15 lm scaling factor = 0.3	word entrance penalty = 4 lm scaling factor = 0.5
64.9 %	64.3 %
speaker stats	speaker stats