Acoustic Adaptation and Accent Identification in the ICSI MR and FAE Corpora

Javier Macías-GuarasaInternational Computer Science Institute

Berkeley, CA - USA

Acoustic Adaptation and Accent Identification in the ICSI MR and FAE Corpora

2

Overview

• Introduction• Acoustic adaptation

– MR SI task– MR SD task

• Accent identification– MR SI task– FAE task

• Conclusions• Future work

3

Introduction (I)

• Work on improving WER for non-native speakers in the ICSI MR corpus

• General details on the Meeting Recording corpus:– Number of speakers: 61– Speech segmented: 85:08:21– Number of accents: 15– ‘Workable’ accents:

• American 53:12:35 15m+8f• German 11:37:01 10m+2f• Spanish 04:38:24 4m+1f• British 01:03:45 2m+0f just for reference

4

Introduction (II)

• Initial idea:– Pronunciation modeling for non native

speakers

• Acoustic adaptation techniques to be tested first:– SRI Decipher system capabilities:

• MAP/MLLR/PhoneLoop

– Analyze different strategies

• Speaker dependent and independent tasks

5

Introduction (III)

• Accent identification:– Needed to effectively use accent-dependent

models in a real-world system– Emphasis in ‘practical’ approaches using,

again, SRI Decipher capabilities

• MR task is a difficult acoustic environment:– Low number of speakers/speech material– Certain speakers dominance (more details?)

• FAE task also approached

8

Introduction (VI)Baseline WERs

• Using SRI 2003 system, WER:

40.3%34.1%

52.3%

104.2%

95.6%

41.4%

33.0%

51.6%

88.2%

65.0%

0%

20%

40%

60%

80%

100%

120%

All American German Spanish British

New SI partition

New SD partition

11

Acoustic adaptation (I)

• Initial studies with old partitioning shows that global task adaptation through MAP is the best approach:– Accent-dependent MAP adaptation also promising

• Initial attempt to do full retraining using 16KHz speech (also 8KHz speech as reference):– Very bad results (more details?)

• Worse than baseline!!

– Too few speakers in the training set given the task partition (speaker independent)

12

Acoustic Adaptation (II) Previous work

• Interest in language learning tools (CALL)

• Standard acoustic adaptation techniques– MAP/MLLR using L1 or L2 speech data– Model interpolation– Clustering– Sufficient for high proficiency speakers

• Pronunciation modeling:– Little (if any) success reported

14

Acoustic Adaptation (IV) Objectives

• Strategies for SI task, ¿combined improvement?:– Task MAP adaptation (TaskMAP)– Accent dependent MAP (AccMAP)– TaskMAP followed by AccMAP/MLLR

• Strategies for SD task:– Task MAP adaptation (TaskMAP) (includes

speaker adaptation)– Per speaker MAP adaptation (SpkMAP)

15

Acoustic adaptation (V)

• Strategies for Acoustic adaptation:– Adaptation weights tuned per accent (heldout)

– Final phoneloop stage

MAP(task adaptation)Full DB MAP/MLLR

SWBmodels

MAP/MLLR

MAP/MLLR

Global MAPmodels

.

.

.

Am DB

Ge DB

Sp DB

Am MAPmodels

Ge MAPmodels

Sp MAPmodels

OR?

TaskMAP

AccMAPTask+AccMAP

17

Acoustic Adaptation (VII) MR Speaker Independent Task

• SI adaptation summary, WER:

34.1%

52.3%

104.2%

95.6%

30.4%

42.3%

93.2%87.9%

0%

20%

40%

60%

80%

100%

120%

American German Spanish British

Baseline SI

TaskMAP-optimal

AccMAP-optimal

TaskMAP + AccMAP-optimal

+ phoneloop pass2

18

Acoustic Adaptation (VIII) MR Speaker Independent Task

• SRI 5xRT system:– Using new dictionary and interpolated LMs– Using best map adapted models for mel

features– Still some bugs in the process (more details?)

American German Spanish BritishBest single system 30.4% 42.3% 93.2% 87.9%

SRI 5xRT system adapted 33.6% 44.9% 86.6% 79.7%SRI 5xRT system STD 31.0% 44.1% 93.4% 78.7%

err (%rel) [ 10.5% ] [ 6.1% ] [ -7.1% ] [ -9.3% ]

20

Acoustic Adaptation (X) MR Speaker Dependent Task

• SD adaptation summary, WER:

33.0%

51.6%

88.2%

65.0%

29.5%

37.1%

54.3%

60.4%

0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

American German Spanish British

Baseline SDTaskMAP-optimalSpkMAP-optimal+ phoneloop pass2

21

Accent Identification (I)• Background:

– Techniques similar to Language Identification– GMM based:

• Broad collection of features• GMM tokenizers

– Broad phonetic classes + HMMs– LM/AM score comparison– Based on phonotactic characteristics:

• PPRLM, PRLM

– More complex than LID– Hard to compare rates: No previous work in MR/FAE

23

Accent Identification (II) Objectives

• Strategy: Use SRI Decipher characteristics– Practical approach: Reasonable run times– GMM classification module (for gender detection)

• Evaluate standard features and normalizations

– Hypothesis driven, phone recognition:• CD/CI models • Recognition using flat Phone LM or flat LM• View as a text classification problem

– Phone LM driven:• PRLM/PRLM

– Combination using NNs

24

Accent Identification (III) MR data: MM classification approach

• GMM results for MR corpus:– Unbalanced data tested over and under

sampling– Use different features & normalization:

• No significant differences (except when using voicing features):

– lack of data– ~Uniform channels

23855 5814 2335 288American German Spanish British ID rate Naive rate err (%rel)

fc downsampling 256 96.1% 65.9% 0.0% 0.0% 82.9% 73.9% -34.5%fc 2048 96.1% 71.6% 0.0% 0.0% 83.9% 73.9% -38.2%

25

Accent Identification (IV) MR data: GMM classification approach

• GMM results for MR corpus :– As a function of utterance length task

AM-GE-SP-BR

60%

65%

70%

75%

80%

85%

90%

95%

100%

1 10 100> Utterance length (in seconds)

AI

rate

Chance

256

26

Accent Identification (V) MR data: Hypothesis driven approach

• Text classification view using MR data:– Input from phone recognition:

• From free phone recognition (CD/CI models, full/flat PLM)

– Rainbow: CMU tool for text classification• Naive bayes classification technique• N-grams (1..6)• No further restrictions (feature selection, stop list,

etc.)

27

Accent Identification (VI) MR data: Hypothesis driven approach

• Text classification view using MR data:– Best results using CI models + flat PLM

(bigrams & trigrams)– Chunk based classification rates (simulation):

Chunks ID rate Naive rate err (%rel)

Full task 84.6% 65.8% -55.0%american nonnat 83.8% 68.0% -49.3%american german 95.9% 81.8% -77.4%american spanish 92.8% 90.2% -26.6%american british - 98.1%am ger bri spa 89.3% 75.0% -57.2%

am ger spa 90.8% 76.1% -61.3%

28

Accent Identification (VII) MR data: Hypothesis driven approach

• Text classification view using MR data:– Utterance based classification rates

(simulation):

– Need longer sequences!!

Utterances ID rate Chunk rate Naive rate err (%rel)

Full task 64.88% 84.6% 66.8% 5.7%american nonnat 65.21% 83.8% 69.1% 12.6%american german 80.16% 95.9% 92.1% 152.4%american spanish 91.79% 92.8% 92.1% 4.5%american british 80.42% - 97.0% 544.1%am ger bri spa 73.19% 89.3% 74.8% 6.3%

am ger spa 74.94% 90.8% 76.6% 7.0%

29

Accent Identification (VIII) MR data: Hypothesis driven approach

• Text classification view using MR data:– Real partition classification rates:

– Worse-than-chance rates if utterance based (pending to do length-dependent AI task)

Chunks RealPartition ID rate ID simul Naive rate err (%rel)

Full taskamerican nonnat 75.46% 83.8% 71.2% -14.9%american german 75.16% 95.9% 73.9% -5.0%american spanish 95.24% 92.8% 92.1% -40.0%american british 91.45% - 92.2% 10.2%am ger bri spa 67.68% 89.3% 64.0% -10.2%

am ger spa 73.01% 90.8% 69.3% -12.0%

30

Accent Identification (IX) Phone LM approach

• PRLM: Phone recognition & LM

PhonerecognizerSpeech

AMLM scoring

LM scoring

PLM accent 1

PLM accent N

Scorecomparison

Scorecomparison

Scorecomparison

Decision...

31

Accent Identification (X) Phone LM approach

• PRLM: Phone recognition & LM– Tested different AMs for phonetic string

generation:• Std forced• Std SWB• MAP adapted per accent• Best is Std SWB

– Tested 1-6gram: • Best is trigram

– But very poor results

32

Accent Identification (XI) MR data: Phone LM approach

• PRLM: Phone recognition & LM:– As a function of utterance length, task

AM-GE-SP-BR: Very bad results

40%

50%

60%

70%

80%

90%

100%

110%

0 10 20 30 40 50 60 70 80 90 100> Utterance length (in seconds)

AI

rate

Chance

StdAM-trigram

33

Accent Identification (XII) Phone LM approach

• PPRLM: Parallel Phone recognition & LM

Phonerecognizer Models Z

Speech

LM scoringAccent z

Avg accent a

Decision

.

.

.

Phonerecognizer Models A

.

.

.

LM scoringAccent a

LM scoringAccent z

.

.

.

LM scoringAccent a

Avg accent aAvg accent a

Scorecomparison

ScorecomparisonAvg accent a

Avg accent aAvg accent z

.

.

.

Scorecomparison

34

Accent Identification (XIII) FAE database

• Experiments with the FAE database:– 4500 speakers: More acoustic context– 20 seconds per speaker– Proficiency is labeled

• Strategy:– Apply standard techniques – Possibly:

• Use FAE-generated models in MR data

35

Accent Identification (XIV) FAE database: GMM classification

• GMM:– Gender independent classification (16-2048)– FAE results in GE-SP task:

– Norm better than CMN. CMN better than plain features– Pending to test GD models

GMM size fc fasf fc fasf fc fasf fasf+ffvf Naive rate128 54.4% 59.2% 59.2% 63.2% 53.6% 64.8% 61.6% 51.6%256 59.2% 59.2% 59.2% 65.6% 58.4% 72.0% 63.2% 51.6%512 51.2% 60.0% 60.8% 60.0% 60.0% 68.0% 65.6% 51.6%1024 52.8% 64.8% 54.4% 63.2% 56.0% 64.4% 61.6% 51.6%

NormNo norm CMN

36

Accent Identification (XV) FAE database: GMM classification

• GMM:– Combining FAE models with MR data:

• Using frame_cepstrum + CMN (GMM 256)

– Combination is possible, but more experiments are needed!!

GS-SP task German Spanish ID rate Naive rate err (%rel)

MR models 81.1% 40.0% 72.3% 66.0% -18.6%FAE models (cmn) 100.0% 21.9% 73.4% 66.0% -21.8%

37

Accent Identification (XVI) FAE database: hypothesis driven

• Text classification view:– FAE results:

– Better than chance but, still, far from useful– Pending to test FAE models in MR data

ID rate Naive rate err (%rel)

GE-SP 58.9% 51.6% -15.0%FR-GE-IT-SP 36.2% 28.9% -10.3%ALL accents 13.2% 9.4% -4.3%13 accents 17.7% 12.2% -6.2%

38

Accent Identification (XVII) FAE database: Phone LM approach

• PRLM/PPRLM:– Pending

• GMM better than text based classification. GE-SP task, for example:– GMM: 72.0% – Text-based: 58.9%

• Results as a function of speech length to be evaluated

39

Conclusions

• Acoustic adaptation is important to face non-native accents:– MAP adaptation provided best results:

• Task adaptation+accent adaptation

– Work on tuning adaptation weights for SD & SI task (magnitude differences)

– Low proficiency speakers need additional improvements

• Non native speech recognition may not be solvable!

40

Conclusions

• Accent identification:– Proved to be more difficult than LID– Different techniques applied:

• GMM techniques and text classification techniques showed promising results

• Standard PRLM strategy didn’t work as expected (score normalization needed?)

– PPRLM to be tested– Integration to be tested

41

Future work

• Finish current experimentation:– Accent identification:

• Test features and normalizations in GMM and phone LM based

• Test acoustic scores ratios• Test LM scores• Test NN based combination

– NonNat speech characterization:• Errors phone/word• Model ‘usage’ distributions

42

Future work

• Pronunciation modeling:– Evaluation of pronunciation variants found in the

SRI SWB dictionary for NonNat speech– Rule based:

• Rules in German (from Silke Goronzy’s work)• Rules in Spanish• ‘Speaking mode’ probability estimation (accent + …)

• Use of new databases (FAE, TED, Fisher)

43

Future work

• A note on work on pronunciation modeling in the MR task:– The MR corpus is not suitable for data-driven

pronunciation modeling:• High error rates for non native speakers & limited number of

them• Rule based methods are to be tested first

– Initial work on evaluating current pronunciation alternatives is needed

– I got relevant rules for initial testing in German and Spanish

44

Thank you!!

• To ICSI and the ICSI Speech Group, with special emphasis to:– Morgan– Andreas– Qifeng, Barry, Adam, Yang, Yan, Dave, Jeremy,

…– Sven & all international visitors– The FrameNet people (Miriam, Michael & Co.)– Staff, specially Lila, María Eugenia and Diane

45

46

MR Partitioning• Speaker independent (SI subtask)

Male Female Train Test

American 36:07:02 14:58:45 33:12:06 17:53:41

51:05:47 15 8 9m + 5f 6m + 3f

German 11:06:43 0:06:56 7:12:09 4:01:30

11:13:39 10 2 6m + 1f 4m + 1f

Spanish 3:05:47 1:12:39 2:46:57 1:31:29

4:18:27 4 1 2m + 1f 2m + 0f

British 1:03:45 0:00:00 0:54:53 0:08:51

1:03:45 2 0 1m + 0f 1m + 0f

47

Full retraining

• Initial attempt to do full retraining using 16KHz speech:– With old partitioning

– Too few speakers in the training set given the task partition (speaker independent)

All American NonNatSWB models 44.1% 34.5% 82.6%Retrain16K+SWBwordnets 53.0% 46.2% 80.1%Retrain16K+SWBwordnets+newHLDA 51.7% 44.6% 79.8%Retrain8K+SWBwordnets 46.8% 40.5% 72.1%Retrain8K+SWBwordnets+newHLDA

48

Speaker dominance

• Few speakers concentrate most speech material: spkID #length

me013 13:53fe008 6:32me011 5:37mn015 4:55me018 4:19mn007 4:16fe016 3:55me010 3:47mn017 3:01

total 50:15

50

Acoustic adaptation (IV)

• SI TaskMAP adaptation, WER:

– Optimal map weight ~proportional to size of accented speech subset

– Bigger improvements in non native accents– Bigger improvements for bigger data size

All NonNat American German Spanish British

Baseline SI 40.3% 64.5% 34.1% 52.3% 104.2% 95.6%

TaskMAP-optimal 37.8% 57.9% 32.5% 46.4% 95.2% 88.5% err (%rel) [ -6.2% ] [ -10.2% ] [ -4.7% ] [ -11.3% ] [ -8.6% ] [ -7.4% ]

51

Acoustic adaptation (IV)

• SI AccMAP adaptation, WER:

– Similar trends than TaskMAP, but no further improvements, except german benefits from task data!

American German Spanish BritishBaseline SI 34.1% 52.3% 104.2% 95.6%

TaskMAP-optimal 32.5% 46.4% 95.2% 88.5% err (%rel) [ -4.7% ] [ -11.3% ] [ -8.6% ] [ -7.4% ]

AccMAP-optimal 32.5% 46.0% 96.8% 91.9% err (%rel) [ -4.9% ] [ -13.6% ] [ -7.8% ] [ -4.2% ]

Optimal Weight 5 40 40 50

52

Acoustic adaptation (IV)

• SI TaskMAP+AccMAP adaptation, WER:

– Small improvements over TaskMAP– Also tested MLLR instead

(taskMAP+AccMLLR), but no improvements

American German Spanish BritishBaseline SI 34.1% 52.3% 104.2% 95.6%

TaskMAP-optimal 32.5% 46.4% 95.2% 88.5% err (%rel) [ -4.7% ] [ -11.3% ] [ -8.6% ] [ -7.4% ]

+ AccMAP-optimal 32.5% 45.8% 95.0% 88.6% err (%rel) [ 0.0% ] [ -1.3% ] [ -0.2% ] [ 0.1% ]

Optimal Weight 5 20 30 40

53

Gender ID issues

• Gender identification:– Per chunk gender ID:

– Per utterance gender ID:

#chunks Male Female ID rateTrue male 350 100.0% 0.0%

True female 97 8.3% 91.8%

#utterances Male Female ID rateTrue male 68304 100.0% 0.0%

True female 19563 2.9% 97.1%

98.2%

99.4%

#utterances Male Female ID rateTrue male 68304 88.5% 11.5%

True female 19563 22.6% 77.4%86.0%

54

Acoustic AdaptationSRI 5xRT System results

British 002-mel 002-mel-expanded 005-plp 005-plp-rescored roverSTD 91.5 89.3 78.2 80.6 78.7Adapted 87.9 84.3 78.8 79.7 79.7BestSimple 87.9------------------------------------------------------------------------------Spanish 002-mel 002-mel-expanded 005-plp 005-plp-rescored roverSTD 97.5 94.2 95.9 95.1 93.4Adapted 88.0 85.1 88.4 88.2 86.6BestSimple 93.2------------------------------------------------------------------------------German 002-mel 002-mel-expanded 005-plp 005-plp-rescored roverSTD 50.7 47.6 44.7 44.5 44.1Adapted 45.8 45.8 44.4 44.5 44.9BestSimple 42.3------------------------------------------------------------------------------American 002-mel 002-mel-expanded 005-plp 005-plp-rescored roverSTD 36.3 33.2 31.2 30.8 31.0Adapted 33.6 34.3 33.3 33.1 33.6BestSimple 30.4