cff-version: 1.2.0
message: "If you use PAVO-Bench or the meta-controller in your research, please cite it as below."
title: "PAVO: Pipeline-Aware Voice Orchestration with Demand-Conditioned Inference Routing"
abstract: >
  PAVO treats the ASR-LLM-TTS voice pipeline as a jointly optimizable inference graph
  with demand-conditioned routing. An 85,041-parameter meta-controller, trained with
  multi-objective PPO in 106 seconds on H100, cuts P95 end-to-end latency by 10.3%
  on 200 LibriSpeech samples. The corresponding mean-latency reduction is significant
  at p = 2x10^-6. It reduces per-turn energy by 71% vs a fixed-cloud baseline on a
  50,000-turn simulated benchmark. Hard-constraint masking on the
  inter-stage coupling structure (validated across n=5,430 measurements on three
  LLM families and two hardware platforms) reduces coherence-failure rate from
  7.1% to 0.9% (7.9x), at +110 ms median latency cost.
type: software
license: MIT
url: "https://github.com/vnmoorthy/pavo-bench"
repository-code: "https://github.com/vnmoorthy/pavo-bench"
repository-artifact: "https://huggingface.co/datasets/vnmoorthy/pavo-bench"
keywords:
  - voice-pipeline
  - asr
  - llm
  - tts
  - inference-routing
  - edge-ai
  - latency
  - reinforcement-learning
  - benchmark
authors:
  - family-names: VeiluKanthaPerumal
    given-names: NarasingaMoorthy
    affiliation: University of Pennsylvania
    email: moorthyv@sas.upenn.edu
  - family-names: Imthathullah
    given-names: Mohammed
    affiliation: Google
preferred-citation:
  type: article
  title: "PAVO: Pipeline-Aware Voice Orchestration with Demand-Conditioned Inference Routing"
  authors:
    - family-names: VeiluKanthaPerumal
      given-names: NarasingaMoorthy
      affiliation: University of Pennsylvania
    - family-names: Imthathullah
      given-names: Mohammed
      affiliation: Google
  journal: "Transactions on Machine Learning Research"
  year: 2026