Modelling the evolution of language for modellers and non-modellers

1
Vowel Systems

• Practical Example

2
Why speech?

• Cross-linguistic data available
• On universals
• On acquisition
• On language change
• This data is relatively uncontroversial
• As opposed to e.g. syntax

3
Speech is easy to model

• It is a physical signal
• We can use existing techniques
• Speech synthesis techniques
• Speech processing techniques
• Even neural processing models
• Results are directly comparable to the real thing

4
The aim of the study

• Explain universals of vowel systems
• Why are do certain (combinations of) vowels occur
  more often than others(acoustic distinctiveness)
• How does the optimisation take place?
• Hypothesis
• Self-organisation in a population under
  constraints of production, perception, learning
  causes optimal systems to emerge
• Model
• Agent-based model
• Imitation games

5
Computational considerations

• Simplification 1
• Agents communicate formants, not complete signals
• Greatly reduces the number of computations
• Perception, production already in terms of
  formants
• Simplification 2
• No meaning (problem phonemes are defined in
  terms of meaning)
• Imitation is used instead of distinguishing
  meaning

6
Architecture

• For vowels
• Realistic productionarticulatory
  synthesiser(Maeda, Valleé)
• Realistic perceptionFormant weighting(Mantakas,
  Schwarz, Boë)
• Learning modelPrototype based associative memory

7
The interactions

• Imitation with categorical perception
• Humans hear speech signals as the nearest phoneme
  in their language (?)
• Correctness of imitation depends not only on the
  signals used, but also on the agents repertoires

Initiator
Imitator
8
Imitation failure
Initiator
Imitator
9
Distributed probabilistic optimization

• Pick an agent from the population
• Pick a signal from this agent
• Modify the signal randomly
• Play imitation games with all other agents in the
  population
• If success of modification is higher than success
  of original vowel, keep the change, otherwise
  revert.
• Disadvantage
• Number of signals per agent is fixed beforehand

10
Reactions to imitation game
F2
Shift Closer
F1
Throw away Vowel
Add Vowel
Merge
11
Measures

• Imitative success
• Energy of vowel systems (Liljencrants Lindblom)
• Size
• Preservation
• Success of imitation between agents from
  populations a number of generations apart
• Only in systems with changing populations
• Realism