PaperlayerView exact revision on arXiv

Source-linked AI summary

Pre-trained Speech Processing Models Contain Human-Like Biases that Propagate to Speech Emotion Recognition

Isaac Slaughter, Craig Greenberg, Reva Schwartz, Aylin Caliskan

arXiv:2310.18877v1cs.CLcs.LGcs.SDeess.AS

TL;DR

The paper introduces SpEAT to assess bias in pretrained speech models, addressing the lack of an intrinsic association test for them. Across models, it finds human-like group-valence associations that also tend to align with downstream SER predictions.

  • Problem

    The paper addresses the lack of an intrinsic association and bias evaluation method for pretrained speech-processing models.

    Source: §1 · p. 1

  • Method

    The authors introduce SpEAT, an intrinsic association test, and use it to examine six bias categories across 16 pretrained speech models.

    Source: §1 · pp. 1–2

  • Results

    66 of 96 SpEAT tests (69%) found that groups positively associated with valence in pretrained models also tended to receive higher-valence predictions from downstream SER models.

    Source: §1 · p. 2

  • Takeaways & Limitations

    SpEAT represents social concepts through speaking styles, showing that speech models can reflect bias based on how and by whom speech is delivered, not only language content.

    Source: §6 · p. 9

    The 120 valence samples came from six younger Caucasian speakers and were rated by 10 younger listeners whose race was unspecified, potentially affecting the results.

    Source: §8 · p. 10

Abstract

from arXiv · show

Previous work has established that a person's demographics and speech style affect how well speech processing models perform for them. But where does this bias come from? In this work, we present the Speech Embedding Association Test (SpEAT), a method for detecting bias in one type of model used for many speech tasks: pre-trained models. The SpEAT is inspired by word embedding association tests in natural language processing, which quantify intrinsic bias in a model's representations of different concepts, such as race or valence (something's pleasantness or unpleasantness) and capture the extent to which a model trained on large-scale socio-cultural data has learned human-like biases. Using the SpEAT, we test for six types of bias in 16 English speech models (including 4 models also trained on multilingual data), which come from the wav2vec 2.0, HuBERT, WavLM, and Whisper model families. We find that 14 or more models reveal positive valence (pleasantness) associations with abled people over disabled people, with European-Americans over African-Americans, with females over males, with U.S. accented speakers over non-U.S. accented speakers, and with younger people over older people. Beyond establishing that pre-trained speech models contain these biases, we also show that they can have real world effects. We compare biases found in pre-trained models to biases in downstream models adapted to the task of Speech Emotion Recognition (SER) and find that in 66 of the 96 tests performed (69%), the group that is more associated with positive valence as indicated by the SpEAT also tends to be predicted as speaking with higher valence by the downstream model. Our work provides evidence that, like text and image-based models, pre-trained speech based-models frequently learn human-like biases. Our work also shows that bias found in pre-trained models can propagate to the downstream task of SER.

1 Introduction

The paper introduces SpEAT to measure social bias in pre-trained speech-model representations, addressing where speech-system biases may originate. It finds human-congruent associations across several social groups and alignment with downstream Speech Emotion Recognition (SER) predictions.

1 Introduction

  • SpEAT is introduced as the first intrinsic association and bias evaluation method for pre-trained speech-processing models.It adapts embedding association tests used in computer vision and natural language processing to speech.Source: §1 · p. 1
  • SpEAT measures associations between social groups and valence, which the authors describe as a primary affect dimension and a strong signal in speech and language.Positive valence is often equated with pleasantness, and negative valence with unpleasantness.Source: §1 · p. 1
  • 14–15 of 16 models show positive-valence associations favoring abled, European-American, female, U.S.-accented, and younger speakers over their comparison groups.The respective counts are 14, 14, 14, 15, and 15.Source: §1 · p. 2
  • In 66 of 96 tests (69%), the group favored by SpEAT also tends to receive higher-valence predictions from downstream SER models.The authors select SER because its valence predictions are conceptually related to SpEAT's valence associations.Source: §1 · p. 2
  • The authors suggest that downstream emotion-aware translation or mental-health systems could treat some groups as more angry or sad.These are potential effects of negative valence associations, not effects directly tested in the paper.Source: §1 · p. 2
  • Increasing the number of stimuli from 2 to 10 can double precision, as SpEAT's standard error decreases sharply with sample size.The paper uses bootstrap estimates of standard error to examine uncertainty in the effect size.Source: §1 · p. 2

2 Background and Related Work

Prior speech-bias research mainly compares task performance across social groups, while embedding association tests assess bias in model representations. Pre-trained speech models can be adapted to downstream tasks using their learned representations.

Pre-Trained Models in Speech Processing

  • Pre-trained speech models learn representations from large, initially unstructured datasets and can be adapted to specific tasks using smaller structured datasets.Adaptation may use fine-tuning or feature extraction, which supplies embeddings to a task-specific model.Source: §2 · pp. 2–3

Bias in Speech Models

  • Prior speech-bias studies compare task performance across groups; pre-trained models have been examined mainly through downstream performance differences across speech styles.The reviewed comparisons cover tasks including speech recognition and speaker verification, and social characteristics including gender, dialect, race, age, city, nationality, and native language.Source: §2 · p. 3

Embedding Association Tests

  • Embedding Association Tests measure associations between two target-concept sets and two attribute-concept sets in representation models.They were adapted from human Implicit Association Tests and have been applied to word, sentence, image, and vision-language representations.Source: §2 · p. 3

3 Methodology

SpEAT compares speech-model embeddings for two target concepts with embeddings for two attribute concepts, using relative cosine similarities and an effect size. Its embedding extraction aggregates information across variable-length inputs and model layers.

3 Methodology

  • SpEAT compares embeddings for two target concepts, such as female and male speakers, against embeddings for two attributes, such as positive and negative valence.Its effect size indicates which target set is relatively more similar to the first attribute than to the second.Source: §3 · p. 3
  • For each target embedding, SpEAT subtracts its mean cosine similarity to negative-attribute embeddings from its mean similarity to positive-attribute embeddings.These differences are aggregated across target stimuli to calculate the effect size d.Source: §3 · pp. 3–4
  • Because speech embeddings vary with input length and models have multiple layers, the method aggregates embeddings across layers before comparing them.For each stimulus, it takes the mean embedding within each layer and sums those layer means; the authors use this strategy across models.Source: §3 · p. 4

4 Data

The study evaluates 16 models from four speech-model families and adapts human bias tests to speech stimuli. Its target and valence stimuli are selected and matched to reduce alternative explanations for measured associations.

Pre-Trained Models and Their Training Data

  • The study tests 16 models: three each from wav2vec 2.0, HuBERT, and WavLM, and seven from Whisper.The first three families use self-supervised learning and speech-only training; Whisper uses weak supervision on speech paired with transcripts, with some models trained multilingual.Source: §4 · pp. 4–5

Target and Attribute Stimuli

  • The study adapts visual and audio human IATs to speech, selecting tests with positive and negative valence and target concepts associated with speaking-style differences.The visual tests are drawn from Nosek et al. (2007), and the audio tests from Mitchell et al. (2011) and Pantos and Perkins (2012).Source: §4 · p. 5
  • Figure 1 reports human-congruent positive-valence associations for ability, race, gender, age, and accent in at least 14 of 16 models.The figure compares individual-model SpEAT results and their mean with human IAT D values.Source: Figure 1 · p. 6
  • Target speech samples are matched across groups on gender and approximate age, and generally on speech content, with specified exceptions.The exceptions include female–male, young–old, human–synthesized, and abled–disabled tests for matching variables; the European-American–African-American test uses extemporaneous speech and is not matched on content.Source: §4 · p. 6
  • Positive and negative valence attributes come from MESS recordings, with the ten highest- and lowest-rated samples selected for each speaker.The resulting attribute sets contain 60 positive and 60 negative samples, balancing speaker representation across valence poles.Source: §4 · pp. 6–7
  • Only one of 16 transcript-sentiment scores differed significantly by race (p = 0.02), and that score favored the group not generally favored by the speech models.The result is from the positive sentiment score of the model from Loureiro et al. (2022).Source: §4 · p. 6

5 Evaluation and Results

SpEAT results largely matched human IAT associations, and its valence associations often aligned with downstream SER predictions. The evaluation also found that estimated uncertainty decreased with more stimuli, alongside a group-specific model-family effect.

Comparison to Human Biases

  • 79 of 96 tests (82%) had positive SpEAT d values, reflecting the same biases recorded in society.The compared human IAT D values were positive and ranged from 0.31 to 0.49.Source: §5 · p. 7
  • Only 6 of 16 models showed bias against synthetic voices, while 14 or more showed stereotype-congruent biases for all other tests.Source: §5 · p. 7
  • The model-family-by-social-group interaction was just below the p=0.05 threshold (p=0.048), but the main effect of family was not significant.Models in the same family tended toward similar SpEAT d values for a given social group, not across all groups.Source: §5 · p. 7

Bias Propagation to Speech Emotion Recognition

  • For the downstream test, each of 16 pre-trained models was adapted to predict valence using the 1,800-sample MESS dataset.The study compared standardized mean predicted valence between matched stimuli from each pair of target groups.Source: §5 · p. 8
  • 66 of 96 tests (69%) aligned: the group more associated with positive valence upstream was also predicted as higher in valence by downstream SER models.Across tests other than human versus synthesized speech, average Cohen’s d and average SpEAT d were in the same direction.Source: §5 · p. 8

Uncertainty of SpEAT d

  • 0.75, 0.33, and 0.12 were the average SEs at 2, 10, and 60 stimuli per target group, respectively.The reported averages decreased as sample size increased; Figure 3 plots SE estimates and their average across bias types and speech models.Source: §5 · p. 8, Figure 3 · p. 8

6 Discussion

The discussion connects human-like associations in speech models to the social signals in training data and to downstream bias when upstream and downstream measures are conceptually related. It also identifies limits to the tested categories and calls for further study of mitigation.

Discussion

  • The authors hypothesize that co-occurrence in training data may contribute to learned associations, such as speech styles being linked with positive valence.They also suggest that dispersion in the human-versus-synthesized test may relate to synthesized speech being absent from the pre-training corpora.Source: §6 · p. 8
  • Representing concepts through speaking styles captures associations based on how speech is delivered and by whom, not just its linguistic content.The discussion notes that data can also reflect who produced or recorded it, whom it concerns, and whom it has meaning for.Source: §6 · p. 9
  • The results add evidence of a relationship between upstream and downstream bias when the two measures concern related concepts.Here, both the SpEAT and the downstream SER comparison concern valence; cited prior studies found mostly uncorrelated or slight relationships across other settings.Source: §6 · p. 9
  • The authors call for detailed study of mitigation; they are unaware of technical solutions specific to implicit bias in pre-trained speech models.They discuss oversampling, adversarial training, and embedding modification as strategies explored in vision and NLP, while noting dataset composition may not be a complete solution.Source: §6 · p. 9

7 Conclusion

The paper introduces SpEAT to measure bias in pre-trained speech models and reports associations across social groups, with evidence of alignment to downstream SER bias. The authors conclude that upstream bias may propagate when the two biases are connected.

7 Conclusion

  • SpEAT is a method for measuring bias in pre-trained speech processing models.Source: §7 · p. 9
  • Across 16 pre-trained models, the study found consistent valence associations involving speakers of different ages, races, genders, and abilities.The study tested six categories of bias.Source: §7 · p. 9
  • Bias detected with SpEAT was correlated with bias in downstream models adapted to SER.The conclusion frames propagation as possible when upstream and downstream biases are connected.Source: §7 · p. 9

8 Limitations

The evaluation covers selected popular models and biases previously documented in humans, but does not capture the full range of biases in speech models. The authors also identify limits in the valence stimuli’s speaker and listener demographics.

8 Limitations

  • The study’s selected social characteristics do not capture the full extent of biases that may be present in speech models.The work focused on popular models and biases in speech styles previously documented in humans using IATs.Source: §8 · p. 10
  • Quantifying biases for a wider range of gender identities or ethnicities will require more consideration and data.Source: §8 · p. 10
  • The 120 valence samples came from six younger Caucasian speakers and were rated by 10 younger listeners; results may be affected by these demographics.The listeners’ race information was not given.Source: §8 · p. 10

9 Ethics Statement

The paper focuses on bias in pre-trained models and its connection to downstream Speech Emotion Recognition (SER), while noting that SER raises other concerns.

9 Ethics Statement

  • The authors describe detecting and measuring bias as a way to raise awareness, avoid harmful impacts, and study bias.Source: §9 · p. 10
  • SER may reinforce norms about emotional expression, enable emotional manipulation, and degrade privacy and autonomy.The authors caution that bias is only one of several potential issues with SER.Source: §9 · p. 10
  • SER can at best capture conveyed emotion or emotion perceived by an annotator, rather than the speaker’s inner state.Source: §9 · p. 10

Disclaimer

The disclaimer states that the results should not be treated as endorsements or official government findings.

Disclaimer

  • The results are not official findings on the part of NIST or the U.S. Government.Source: Disclaimer · p. 10
  • The results should not be construed as an endorsement of any participant’s system or methods.Source: Disclaimer · p. 10
  • The results should not be construed as an endorsement of any participant’s commercial product.Source: Disclaimer · p. 10

A Statistical Significance of SpEATs

The appendix reports significance testing for SpEAT results and provides details on evaluated models, target stimuli, and embedding aggregation. It also describes how the tests and stimuli were selected across model families and social-group comparisons.

A Statistical Significance of SpEATs

  • 63 of 96 SpEAT tests (66%) were significant at α = 0.01, and 46 of those 63 (73%) remained significant after Bonferroni correction.The correction addresses the increased chance of at least one accidental rejection across many simultaneous tests.Source: A Statistical Significance of SpEATs · p. 15, Figure 4 · p. 16

C.1 Abled and Disabled Speakers

  • For abled-versus-disabled speech, the study uses dysarthric and non-dysarthric speakers, chiefly from TORGO, and leaves other speech-manifesting disabilities for future work.TORGO samples were matched by gender and speech content, but not age; UASpeech provides an additional dataset test using words or phrases.Source: §C.1 · pp. 17–18

C.2 European-American Speakers and African-American Speakers

  • The European-American versus African-American comparison uses Buckeye and CORAAL interview speech, with samples matched or balanced by speaker age and gender.The sampled groups each contribute 60 samples, and their clips did not differ in average length (t(59) = −0.12, p = 0.90).Source: §C.2 · p. 18

C.3 Female and Male Speakers

  • The female-versus-male test uses age-matched U.S.-born English speakers from the Speech Accent Archive; Appendix E also tests the original human-IAT stimuli.Those original stimuli include human speakers and synthesized voices producing the same short neutral phrases.Source: §C.3 · p. 19

C.4 Human Speakers and Synthesized Speech

  • The human-versus-synthesized speech test uses SAA speech and synthetic voices based on systems similar to those in the original human IAT.Synthetic samples vary speed and are matched to SAA samples by gender; Appendix E also evaluates the original stimuli.Source: §C.4 · p. 19

C.5 United States and Foreign Accented Speakers

  • The U.S.-versus-Korean-accent comparison uses age- and gender-matched SAA speakers, with a separate validation using the original IAT stimuli.The original stimuli use two similarly aged male actors reading the same neutral phrases in their native accents.Source: §C.5 · pp. 19–20

C.6 Young and Old Speakers

  • The young-versus-old comparison uses gender-balanced U.S.-born speakers aged 18–19 and 57–93, with 58 speakers per target group.Source: §C.6 · p. 20

D Other Embedding Aggregation Strategies

  • Alternative embedding aggregation strategies mostly produced positive SpEAT d values, with mean temporal aggregation and early-to-middle layers tending to align more with human biases.Table 2 describes temporal aggregation across variable-length sequences and layer aggregation across transformer layers.Source: D Other Embedding Aggregation Strategies · p. 20, Table 2 · p. 20

E SpEAT Results Using Alternate Stimuli Sets

Three of four alternate-stimuli tests show means similar to the original tests, but one U.S.-versus-foreign speech test reverses direction. The paper also describes downstream SER models and tests using U.K. English.

E SpEAT Results Using Alternate Stimuli Sets

  • Three of four alternate-stimuli tests have means similar to the original tests; the U.S.-versus-foreign test using Pantos and Perkins stimuli instead reverses direction.The authors hypothesize that its eight stimuli, two speakers, and repeated words or phrases may help explain the difference.Source: E SpEAT Results Using Alternate… · p. 20, Figure 5 · p. 21

F Downstream Speech Emotion Recognition Models

  • Downstream SER models weight pre-trained model layers, project embeddings to 256 dimensions, average across the audio sequence, and predict a single valence score.The models use mean squared error and are trained at learning rates of 10^-3, 10^-4, and 10^-5.Source: F Downstream Speech Emotion… · p. 21

G Embedding Association Tests for U.K. English

G.1.1 Target Stimuli

  • The U.K. accent test uses 32 samples per group from two female actors, while the age test uses 12 speakers per group from the SAA and balances gender.The age groups are 18–21 and 38–72; the authors describe a tradeoff between larger groups for precision and smaller groups for better category representation.Source: §G.1.1 · pp. 21–22

G.1.2 Attribute Stimuli: EU-Emotion Stimulus Set

  • For U.K. valence stimuli, the authors use EUESS semantically neutral sentences and sample positive and negative recordings within speaker.Human ratings showed no evidence of gender differences in valence (t(437) = 0.81, p = 0.42) or a relationship between age and valence (t(438) = −0.90, p = 0.37).Source: §G.1.2 · p. 22, Figure 7 · p. 23

G.2 Test Results

10 of 16 models showed a positive SpEAT d value in the British- and foreign-accented-speaker test, while only two models showed bias against older U.K. speakers.

G.2.1 British and Foreign Accented Speakers

  • 10 of 16 models show a positive SpEAT d value in the British- and foreign-accented-speaker test.The samples came from only 16 recordings by two speakers, and their British locality may have affected the results.Source: §G.2.1 · p. 23
  • The British- and foreign-accented samples may not represent their intended concepts because each group was represented by only two speakers and 16 recordings.The British locality of the speech may also affect the results.Source: §G.2.1 · p. 23

G.2.2 Young and Old Speakers

  • Only two models showed bias against older U.K. speakers; the test used 12 samples per concept, and its older-speaker age cutoff was lower than the U.S. cutoff.The lower cutoff may make the U.K. target concepts less differentiated.Source: §G.2.2 · p. 23
Loading 2310.18877v1…