[Newman, Clouse, & Burnham, 2001]

[Newman, Clouse, & Burnham, 2001]

An unfamiliar talker appears!

Listeners have to simultaneously infer what the talker is saying and how they say things

An ideal adpater rapidly adapts to an unfamiliar talker via distributional learning

[Kleinschmidt & Jaeger, 2015]

Distributional learning needs to start somewhere

(prior beliefs in Bayesian jargon)

Scylla and Charybdis

Too much uncertainty

Overconfident and wrong

Ideal adapter predicts:

• Listeners’ expectations about an unfamiliar talkers should be as specific as possible while still being right on average.
• Adaptation is constrained by actual variability across talkers (as experienced by each listener).
• Accents that fall outside the range of normal variability should result in less adaptation.

Questions

1. Is rapid adaptation to an unfamiliar talker constrained?
2. Are constraints consistent with belief updating, starting from shared prior beliefs?
3. Are these inferred prior beliefs ‘rational’ in that they reflect cross-talker variation?

VOT drawn from a bimodal distribution

Distributional learning

Design

• Five “accents” with different VOT distributions
• \(n=138\) subjects on Mechanical Turk
• 222 trials (about 20 minutes)

Prediction: no learning

Prediction: full learning

Results: classification

Results: category boundaries

Questions

1. Is rapid adaptation to an unfamiliar talker constrained?
   • Yes: Less adaptation to extreme accents
2. Are constraints consistent with belief updating, starting from shared prior beliefs?
3. Are these inferred prior beliefs ‘rational’ in that they reflect cross-talker variation?

Belief updating

Previously: given exposure distributions and prior beliefs, predict adaptation (change in classification)

Now: given adaptation (change in classification) to different exposure distributions, infer prior beliefs.

[Kleinschmidt & Jaeger, 2015]

Questions

1. Is rapid adaptation to an unfamiliar talker constrained?
   • Yes: Less adaptation to extreme accents
2. Are constraints consistent with belief updating, starting from shared prior beliefs?
3. Are these inferred prior beliefs ‘rational’ in that they reflect cross-talker variation?

Model vs. data: classification

Model vs. data: classification

Questions

1. Is rapid adaptation to an unfamiliar talker constrained?
   • Yes: Less adaptation to extreme accents
2. Are constraints consistent with belief updating, starting from shared prior beliefs?
   • Yes: Belief updating model fits classification well
3. Are these inferred prior beliefs ‘rational’ in that they reflect cross-talker variation?

Inferred prior beliefs

Inferred prior beliefs

Inferred prior beliefs

Questions

1. Is rapid adaptation to an unfamiliar talker constrained?
   • Yes: Less adaptation to extreme accents
2. Are constraints consistent with belief updating, starting from shared prior beliefs?
   • Yes: Belief updating model fits classification well
3. Are these inferred prior beliefs ‘rational’ in that they reflect cross-talker variation?
   • Yes(-ish): Prevoicing potentially explains low expected /b/ VOT

Distributional learning is constrained

• Things that can’t be the whole story to coping with talker variation:
  • Rapid adaptation
  • Fine-grained episodic memory
  • Normalization

There are other possible sources of constraints!

• Psychoacoustics
• Own production processes
• etc.

The ideal adapter

• Provides a theory for why and how the statistics of prior experience guide (and constrain) adaptation
• …and supports extraction of social information from the speech signal

Other contrasts and groups

• Ideal adapter predictions depend on type and amount of cross-talker variability.
  • (relatively) weaker constraints for more variable contrasts (e.g., fricatives and vowels)
  • (relatively) stronger constraints for more specific groups (e.g., gender and dialect)

Mind reading

• Probing listeners’ subjective expectations is hard.
• Adaptation + belief updating models provide a missing tool.
• Doesn’t require production data.

Findings

1. Rapid adaptation via distributional learning is constrained
2. Proof of concept: recover listeners prior beliefs about unfamiliar talkers
3. Provides a new tool that links (sociolinguistic) variation in production with listeners’ perception expectations

Does anyone actually prevoice

[Dmitrieva et al., 2015, doi:10.1016/j.wocn.2014.12.005]

Belief updating