PAVO-Bench: 50K-Turn Benchmark for ASR-LLM-TTS Pipeline Routing

Code: github.com/vnmoorthy/pavo-bench · Paper: TMLR 2026 (accepted) · Authors: NarasingaMoorthy VeiluKanthaPerumal (UPenn), Mohammed Imthathullah (Google)

pip install git+https://github.com/vnmoorthy/pavo-bench.git

Headline results (vs fixed-cloud baseline, 50,000 voice turns)

Metric	Result	Significance
P95 end-to-end latency (H100, LibriSpeech)	−10.3% (−167 ms)	—
Median latency	−34%
Energy per turn	−71%
Coherence-failure rate	7.1% → 0.9% (7.9× reduction)	hard-constraint masking, +110 ms median cost
Meta-controller size	85,041 parameters	—
Meta-controller training	106 seconds on H100	—

The p = 2×10⁻⁶ result applies to the paired test of mean end-to-end latency (2,277 vs 2,671 ms over five bootstrap replications), not the descriptive P95 comparison; the paired Wilcoxon test on the same five replications gives p = 0.0625.

The empirical contribution is a two-regime coupling structure (sharp factual-accuracy cliff + gradual semantic degradation) characterized over n = 5,430 measurements across two hardware platforms (H100, Apple M3) and three LLM families (Llama 3.1 8B, Mistral 7B, Gemma2 2B).

Description

PAVO-Bench evaluates ASR-LLM-TTS voice pipeline routing decisions. It provides 50,000 turns of benchmark data designed to measure how well different pipeline configurations balance latency, quality, cost, and energy when routing spoken-language queries through cascaded ASR, LLM, and TTS components.

The benchmark is organized into three tiers plus component-level ablation. The 50K routing evaluation is simulated using measured H100/M3 latencies and published benchmark values for unavailable configurations; the repository also includes direct-inference H100 results.

Dataset Files

The Hugging Face default config exposes the 40K/10K JSONL files as standard train and test splits. The heterogeneous result-summary JSON files below remain directly downloadable artifacts rather than dataset splits, which keeps the Dataset Viewer schema valid.

Primary dataset

File	Split	Rows
results/tier3_50k_train.jsonl	train	40,000
results/tier3_50k_test.jsonl	test	10,000

Tier 1 — Unit-Level Validation

File	Description
tier1_statistical_results.json	Statistical reproducibility across 5 trials × 1,000 turns each (seeds 42, 123, 456, 789, 1024).
tier1_coupling_results.json	Coupling-cliff calibration — LLM quality degradation vs ASR word-error rate (WER 0–20%).
tier1_llm_latency_results.json	Latency profile for llama3.1:8b across short / medium / long generation contexts.

Tier 2 — Integration-Level Evaluation

File	Description
tier2_e2e_results.json	End-to-end pipeline measurements (cloud_premium, ondevice_fast, hybrid_balanced, pavo_adaptive) on 200 LibriSpeech samples.
tier2_cross_dataset_results.json	Cross-dataset ASR (LibriSpeech + FLEURS) for whisper-large-v3 and whisper-tiny.
tier2_noise_robustness_results.json	ASR robustness at SNR 5–30 dB plus clean baseline.

Tier 3 — Scale Evaluation

File	Description
tier3_50k_summary.json	Summary statistics for the 50K-turn dataset (40K train / 10K test split, complexity 1–5).
tier3_scaling_results.json	Per-model latency benchmarks for simple / medium / complex queries.

Component Analysis

File	Description
component_ablation_results.json	PAVO-Full vs PAVO-NoCoupling, Always-Cloud, Always-OnDevice, etc.

Usage

from huggingface_hub import hf_hub_download
import json

path = hf_hub_download(
    repo_id="vnmoorthy/pavo-bench",
    filename="tier3_50k_summary.json",
    repo_type="dataset",
)
print(json.load(open(path)))

Or via the pip package:

from pavo_bench import load_dataset, PretrainedPAVORouter, benchmark_router
turns = load_dataset(split="test")
pavo  = PretrainedPAVORouter.from_released()
print(benchmark_router(pavo, turns))

Or directly with datasets:

from datasets import load_dataset

dataset = load_dataset("vnmoorthy/pavo-bench")
print(dataset["train"].num_rows, dataset["test"].num_rows)  # 40000 10000

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}
}

License

Dataset, results, coupling matrices, and model weights: CC-BY 4.0. Code is MIT licensed in the linked GitHub repository.