APEX-Voice

Measuring professional work completion by full-duplex voice agents.

APEX-Voice is a benchmark of 120 professional tasks for realtime speech-to-speech voice agents. In each task, a full-duplex agent talks to a simulated user who streams real speech, with overlap, back-channels, barge-ins and mid-speech corrections. While talking, the agent calls native tools in an instrumented environment to produce a professional artifact, such as a benefits enrollment, an insurance claim, a negotiation record or an inspection report. The benchmark grades the artifact and the environment state, not the conversation.

• Code, harness and model adapters: https://github.com/apex-voice/code
• Scored runs behind the paper (1,800 sessions + judge cache): puneetUMD/APEX-Voice-Runs
• License: Apache 2.0. All people, companies, identifiers and records are synthetic.

Quick start

pip install "apex-voice[all] @ git+https://github.com/apex-voice/code"
apex-voice download --local-dir APEX-Voice
apex-voice validate --data APEX-Voice               # offline: 120/120 task packages valid
apex-voice run --model gpt-realtime --data APEX-Voice   # replays the pre-rendered user audio

To browse the task index (one row per task) with 🤗 Datasets:

from datasets import load_dataset

index = load_dataset("puneetUMD/APEX-Voice", split="test")
print(index[0]["title"], index[0]["archetype"], index[0]["primary_artifact"])

The index is for browsing and filtering. To run the benchmark you need the full task packages under tasks/, which apex-voice download (or huggingface_hub.snapshot_download) fetches.

Dataset structure

tasks/apexv1_001 … apexv1_120/    # one self-contained task package per task
tasks/apexv1_*/user/audio/        # pre-rendered simulated-user speech (FLAC) + index.jsonl
data/tasks.jsonl                  # flat index: one row per task (the viewer / `load_dataset` split)
SHA256SUMS                        # checksums of every file under tasks/ and data/

Each task package contains:

file	contents
task.yaml	id, title, profession, industry, archetype, risk tier, time budget, autonomy level, actions requiring approval
taxonomy.yaml	labels on every taxonomy axis
initial_state.json, workspace/*.json	initial environment state, initial workspace, latent world
tools/tool_spec.yaml	native function tools exposed to the agent (read / draft / commit)
knowledge/manifest.json	knowledge-base documents (gold and distractors)
grading/gold.yaml	hidden grader: terminal state, required and forbidden actions, critical gates, per-field artifact expectations
reference/policy.yaml	oracle trajectory that achieves a perfect score
user/*.yaml, user/realization_bank.jsonl	simulated user: hidden state, flow state machine, act observer, duplex events, persona, and frozen utterance realizations
user/audio/	pre-rendered simulated-user speech: one clip per utterance variant, plus the fixed sign-off and nudge lines (see below)
assets/inputs/artifact_source_manifest.yaml	artifact provenance

Simulated-user audio (user/audio/)

The simulated user speaks the frozen texts in user/realization_bank.jsonl using Kokoro-82M (snapshot f3ff3571791e39611d31c381e3a41a3af07b4987) in the task persona's voice. Every text the harness can speak is shipped pre-rendered, so a run needs neither torch nor Kokoro and every machine sends the model exactly the same audio:

file	contents
<plan>_<variant>.flac	one clip per realization-bank variant (24 kHz mono, lossless 16-bit)
_closing_<i>.flac, _nudge.flac	the harness's fixed sign-off and "go on?" lines, in the task's voice
index.jsonl	one row per clip: file, kind, user_plan_id, variant_id, text, voice, num_samples, pcm16_sha256, word_timestamps
meta.json	renderer versions (Kokoro revision, kokoro, misaki, torch)

The FLAC samples are exactly the 16-bit PCM the harness streams to the model. apex-voice run replays them by default (--user-audio auto); pass --user-audio kokoro to synthesize live instead. Live Kokoro output is bit-exact only under the same torch build and thread configuration, which is why the clips are published. They match the user channel of the published rep0 and rep1 sessions sample for sample. The published rep2 sessions were synthesized with a different CPU thread configuration: same text, voice and length, with float-level differences (correlation

0.9999) that are inaudible.

Index fields (data/tasks.jsonl)

task_id, version, title, profession, industry, economic_function, archetype, delegation_patterns, user_profile, temporal_dynamics, primary_artifact, autonomy_level, approval_required_for, knowledge_burden, tool_burden, risk_tier, required_fact_count, time_budget_s, tools, duplex_event_types, required_fields, oracle_tool_calls, num_user_plans, path.

Composition

Tasks	120 (99 distinct professions)
Work archetypes	NEGOTIATE 20, COORDINATE 20; FORM_FILL, INTERVIEW, DISCOVERY, INTAKE, TROUBLESHOOT, ADVISE, FACILITATE, INSPECT 10 each
Industries	software/SaaS 43, horizontal enterprise 30, manufacturing & field ops 25, professional services 10, workplace/HR 8, healthcare 3, insurance 1
Artifact classes	11 (negotiation record 20; 10 each for the rest)
Autonomy	A0 prepare-only 51, A1 draft-confirm 29, A2 low-risk execute 16, A3 approval-gated commit 24
Knowledge burden	K0 none 44, K1 supplied 6, K2 search 51, K3 multi-document policy 19
Tool burden	T1 light 44, T2 moderate 76
User profiles	cooperative 24, correction-prone 24, and 12 each of six other profiles
Risk tier	R0 routine 63, R1 sensitive data 25, R2 consequential action 22, R3 special review 10
Required artifact fields	9–14 per task (1,237 total)
Oracle tool calls	11–17 per task (median 13)
Duplex events	181 mid-speech corrections, 2 back-channels
Delegation	every task: delegate / complete / revise; 29 tasks also require explicit user approval

Scoring

A task passes under the Production Task Score (PTS) only if all four gates hold:

• GS: goal satisfied.
• PC: process compliance, e.g. no commit without approval and no stale field after a correction.
• RA: required actions complete.
• WA: every required artifact field is correct and not stale, and the artifact reached its required lifecycle state.

Artifact Field Accuracy (AFA) gives partial credit per field. Fields are graded deterministically first, then by a cached semantic judge. Each task is run 3 times, and results are reported as pass@1, pass@3 and Reliable@3 (the task passes in all 3 runs). Full definitions are in the code repository's docs/metrics.md.

Intended use and limitations

• Intended use. Evaluating realtime voice agents on multi-turn, tool-using professional workflows. The dataset is a test set and is not meant for training.
• Synthetic content. Tasks, people, organizations, identifiers and documents are synthetic. They cover realistic workflows but not every real-world variation of a profession.
• Language. English only. The user speaks with a single TTS system (Kokoro-82M, three voices), so acoustic diversity is limited compared with real callers.
• Judge dependence. Some field verdicts rely on an LLM judge (gpt-4o-mini). Verdicts are cached and versioned so they can be reproduced, but results obtained with another judge model are not directly comparable.
• Contamination. Please do not train on the task packages, especially grading/ and reference/. Doing so invalidates results.

Citation

@misc{mathur2026apexvoice,
  title  = {{APEX-Voice}: Can Voice Agents Complete Professional Workflows Through Full-Duplex Interaction},
  author = {Mathur, Puneet; Dinesh Manocha},
  year   = {2026},
  url    = {https://apex-voice.github.io/}
}

License

Apache License 2.0 (see LICENSE).