pavo-bench

PAVO-Bench: a 50,000-turn voice pipeline benchmark and an 85,041-parameter meta-controller (PPO-trained in 106 seconds) for routing voice-agent calls through cascaded ASR → LLM → TTS pipelines.

• Repo: https://github.com/vnmoorthy/pavo-bench
• Dataset: https://huggingface.co/datasets/vnmoorthy/pavo-bench
• Paper: PAVO: Pipeline-Aware Voice Orchestration with Demand-Conditioned Inference Routing (under review at TMLR)

Install

pip install pavo-bench

To also install Hugging Face dataset support:

pip install pavo-bench[hf]

Quick start

import pavo_bench

# Bundled headline result tables (small JSON summaries):
print(pavo_bench.list_results())
print(pavo_bench.load_results("tier3_50k_summary"))

# Full 50K-turn dataset (requires `pavo-bench[hf]`):
ds = pavo_bench.load_dataset()
print(ds)

Headline numbers

• −10.3% P95 latency on H100, measured directly on 200 LibriSpeech samples (p = 2×10⁻⁶)
• −34% median latency, −71% energy per turn vs. fixed-cloud (50K-turn routing simulation, parameterized by measured per-stage latencies)
• Coupling cliff: Gemma2 2B mean quality drops from 0.825 → 0.585 as ASR WER crosses 2%
• Meta-controller: 85,041 parameters, 3-layer MLP, multi-objective PPO, 106 s training time

License

Code: MIT. Dataset: CC-BY 4.0.

Citation

@article{veilukanthaperumal2026pavo,
  title   = {PAVO: Pipeline-Aware Voice Orchestration with Demand-Conditioned Inference Routing},
  author  = {VeiluKanthaPerumal, NarasingaMoorthy and Imthathullah, Mohammed},
  journal = {Transactions on Machine Learning Research},
  year    = {2026}
}