VLBR Coding in Noisy Environments

Clean Speech Results

Recognition rates under Subway, Car, and Babble noise, extract from paper:

M. Padellini, F. Capman, G. Baudoin, “Very low bit rate speech coding in Noisy Environments”,

submitted to INTERSPEECH’05, Lisbon.

Subway	LPCC			MFCC			AURORA MFCC			MMSE MFCC			PMC LPCC
SNR(dB)	5	15	20	5	15	20	5	15	20	5	15	20	5	15	20
Corr (%)	15	36	45	19	41	47	34	53	57	29	51	52	35	51	62
Sub (%)	82	59	50	76	53	46	59	39	36	65	42	41	41	36	29
Ins (%)	26	28	23	29	23	22	27	21	21	26	22	21	5	10	10
Car	LPCC			MFCC			AURORA MFCC			MMSE MFCC			PMC LPCC
SNR(dB)	5	15	20	5	15	20	5	15	20	5	15	20	5	15	20
Corr (%)	31	57	61	30	55	58	45	61	64	41	53	55	51	72	75
Sub (%)	64	39	34	64	39	36	48	32	29	52	40	38	31	18	16
Ins (%)	34	22	21	26	22	21	22	17	17	26	22	22	6	5	5
Babble	LPCC			MFCC			AURORA MFCC			MMSE MFCC			PMC LPCC
SNR(dB)	5	15	20	5	15	20	5	15	20	5	15	20	5	15	20
Corr (%)	29	54	59	28	51	57	36	54	59	33	50	54	45	63	72
Sub (%)	68	43	38	69	44	39	58	39	36	61	44	40	45	29	22
Ins (%)	38	29	27	37	30	30	33	23	23	30	27	27	19	14	11

Table 1, 2, 3: Recognition rates under Subway, Car and Babble noise.

(Corr: Correct, Sub: Substitution, Ins: Insertion)

Example of noise corrupted files coded with the VLBR coder.

• Features: LPCC parameters with PMC Log Add adaptation of the models.

The noise model has been trained using the first second of the noise corrupted signal, preceding the utterance.

• Unit selection process: adapted by adding noise to the target units.

The noise added is the mean harmonic profile of the first second of the noise corrupted signal, preceding the utterance.

at SNR 5dB problems occur because of the lack of robustness in the pitch estimation.

Clean speech original file: Play,

VLBR coded speech file: Play

	Subway		Car		Babble
	Original	VLBR	Original	VLBR	Original	VLBR
20dB	Play	Play	Play	Play	Play	Play
15dB	Play	Play	Play	Play	Play	Play
5dB	Play	Play	Play	Play	Play	Play