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

Author: NarasingaMoorthy VeiluKanthaPerumal, University of Pennsylvania

Description

PAVO-Bench is a comprehensive benchmark suite for evaluating 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 consumption when routing spoken-language queries through cascaded Automatic Speech Recognition (ASR), Large Language Model (LLM), and Text-to-Speech (TTS) components.

The benchmark is organized into three tiers of increasing scale and complexity, plus component-level ablation studies. All results were produced on GPU hardware.

Dataset Files

Tier 1 -- Unit-Level Validation

File	Description
tier1_statistical_results.json	Statistical reproducibility results across 5 trials of 1,000 turns each (seeds 42, 123, 456, 789, 1024). Reports mean, std, and 95% confidence intervals for PAVO latency, quality, cost, and energy metrics.
tier1_coupling_results.json	Coupling constraint validation measuring LLM quality degradation as a function of ASR word-error rate (WER 0--20%) using llama3.1:8b.
tier1_llm_latency_results.json	LLM latency profiling for llama3.1:8b across short (50 token), medium (200 token), and long (500 token) generation contexts. Reports total latency, time-to-first-token, and tokens/second.

Tier 2 -- Integration-Level Evaluation

File	Description
tier2_e2e_results.json	End-to-end pipeline measurements for cloud_premium (whisper-large-v3 + llama3.1:8b) and edge_fast (whisper-tiny + gemma2:2b) configurations on 200 LibriSpeech samples. Includes per-stage latency breakdowns, sample ASR outputs, and sample LLM responses.
tier2_cross_dataset_results.json	Cross-dataset ASR evaluation on LibriSpeech and FLEURS for whisper-large-v3 and whisper-tiny models (200 samples each). Reports WER and latency statistics.
tier2_noise_robustness_results.json	ASR robustness under white noise at SNR levels 5--30 dB, plus clean baseline. Reports WER degradation across noise conditions.

Tier 3 -- Scale Evaluation

File	Description
tier3_50k_summary.json	Summary statistics for the full 50,000-turn PAVO-Bench dataset: 40K train / 10K test split, complexity distribution (levels 1--5), generation time, and error rate.
tier3_scaling_results.json	LLM scaling benchmarks across multiple models (gemma2:2b, llama3.1:8b, etc.) with simple, medium, and complex query types. Reports latency, throughput, and real-time suitability.

Component Analysis

File	Description
component_ablation_results.json	Ablation study comparing PAVO-Full, PAVO-NoCoupling, and other ablated configurations. Reports latency, quality, cost, energy, coupling violations, and infeasible percentages.

Usage

Load individual JSON files directly

import json
from huggingface_hub import hf_hub_download

# Download a specific results file
path = hf_hub_download(
    repo_id="<your-username>/pavo-bench",
    filename="tier3_50k_summary.json",
    repo_type="dataset",
)

with open(path) as f:
    data = json.load(f)

print(f"Total samples: {data['total_samples']}")
print(f"Train/Test split: {data['train_samples']}/{data['test_samples']}")

Download all files

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="<your-username>/pavo-bench",
    repo_type="dataset",
    local_dir="./pavo-bench-data",
)

Benchmark Metrics

• Latency (ms): End-to-end and per-component response time
• Quality (0--1): Composite score incorporating ASR accuracy and LLM response quality
• Cost (USD): Per-turn inference cost
• Energy (mJ): Per-turn energy consumption
• Coupling violations: Cases where ASR errors propagate and degrade LLM quality

Citation

If you use PAVO-Bench in your research, please cite:

@misc{pavo-bench-2026,
  author = {VeiluKanthaPerumal, NarasingaMoorthy},
  title = {PAVO-Bench: A 50K-Turn Benchmark for ASR-LLM-TTS Pipeline Routing},
  year = {2026},
  institution = {University of Pennsylvania},
  url = {https://huggingface.co/datasets/<your-username>/pavo-bench}
}

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC-BY 4.0) license.