Voice Agent Evaluation Benchmarks
Collection
Public benchmarks for voice agents, speech tools, computer action, ASR, and meeting understanding, with a structured landscape. • 14 items • Updated
id stringlengths 3 34 | name stringlengths 3 34 | primary_focus stringlengths 33 100 | audio_input stringclasses 2
values | audio_output stringclasses 3
values | multi_turn stringclasses 3
values | tool_use stringclasses 2
values | goal_completion stringclasses 3
values | computer_or_browser_action stringclasses 2
values | meeting_or_long_form stringclasses 2
values | real_time stringclasses 3
values | public_data stringclasses 2
values | code_license stringclasses 5
values | data_license stringclasses 7
values | github_url stringlengths 33 58 ⌀ | huggingface_url stringclasses 7
values | paper_url stringclasses 8
values | last_verified timestamp[s]date 2026-09-01 00:00:00 2026-09-01 00:00:00 | evidence_notes stringlengths 136 331 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
voiceagentbench | VoiceAgentBench | End-to-end speech-based agents on realistic tool-driven tasks | yes | no | yes | yes | partial | no | no | no | yes | Krutrim Community License Agreement 1.0 | other | https://github.com/ola-krutrim/VoiceAgentBench | https://huggingface.co/datasets/krutrim-ai-labs/VoiceAgentBench | https://arxiv.org/abs/2510.07978 | 2026-09-01T00:00:00 | Public code and audio-backed data cover single, parallel, sequential, multi-turn, and safety tool-call subsets. Scoring emphasizes tool arguments and refusals rather than live spoken output or computer control. |
audio2tool | Audio2Tool | Spoken tool-call selection and argument extraction across smart-home, wearable, and vehicle tasks | yes | no | partial | yes | no | no | no | no | yes | not stated | CC-BY-NC-4.0 | null | https://huggingface.co/datasets/RVtech/Audio2Tool | https://arxiv.org/abs/2604.22821 | 2026-09-01T00:00:00 | The public dataset contains 16,843 queries and 36,421 audio files across eight tiers, including multi-intent, correction, multi-turn, and overlapping-speech cases. It evaluates audio-to-tool-call generation, not execution against a stateful environment or spoken-response quality. |
eva | EVA | End-to-end conversational voice-agent accuracy and experience | yes | yes | yes | yes | yes | no | no | yes | yes | MIT | MIT | https://github.com/ServiceNow/eva | https://huggingface.co/datasets/ServiceNow-AI/eva-bench | https://arxiv.org/abs/2605.13841 | 2026-09-01T00:00:00 | Bot-to-bot evaluations cover 213 scenarios across three enterprise domains, complete multi-turn spoken conversations, task accuracy, interaction experience, voice perturbations, and 12 evaluated systems. |
voiceassistant-eval | VoiceAssistant-Eval | Listening, speaking, viewing, multi-turn behavior, and safety for general voice assistants | yes | yes | yes | no | no | no | no | no | yes | not stated | MIT | https://github.com/mathllm/VoiceAssistant-Eval | https://huggingface.co/datasets/MathLLMs/VoiceAssistant-Eval | https://arxiv.org/abs/2509.22651 | 2026-09-01T00:00:00 | A broad multimodal voice-assistant suite with public examples across listening, speaking, viewing, multi-turn interaction, and safety. It does not primarily evaluate tool execution or real-world task completion. |
voicebench | VoiceBench | Multi-faceted instruction following, knowledge, reasoning, and safety for LLM-based voice assistants | yes | no | partial | no | no | no | no | no | yes | Apache-2.0 | Apache-2.0 | https://github.com/MatthewCYM/VoiceBench | https://huggingface.co/datasets/hlt-lab/voicebench | https://arxiv.org/abs/2410.17196 | 2026-09-01T00:00:00 | Public audio subsets cover open-ended and multiple-choice QA, instruction following, reasoning, safety, human-recorded speech, and a 46-example multi-turn subset. Evaluation consumes spoken prompts but scores assistant response content rather than spoken-output quality or tool execution. |
voicecomputerbench-talkact | VoiceComputerBench / TalkAct | Real-time voice conversation while operating browser-based computer tasks | yes | yes | yes | yes | yes | yes | no | yes | yes | MIT | MIT | https://github.com/19PINE-AI/TalkAct | null | https://github.com/19PINE-AI/TalkAct/blob/main/paper/paper_arxiv.pdf | 2026-09-01T00:00:00 | VoiceComputerBench evaluates a simulated phone caller and an agent that talks while acting through Playwright on hermetic browser sites. |
tau2-bench | tau2-bench / tau-Voice | Tool-agent-user interaction in realistic domains | yes | yes | yes | yes | yes | no | no | yes | yes | MIT | MIT | https://github.com/sierra-research/tau2-bench | null | https://arxiv.org/abs/2603.13686 | 2026-09-01T00:00:00 | The public framework covers tool use, dynamic user interaction, domain task success, and a full-duplex voice mode. It does not primarily evaluate general desktop or browser control. |
nemo-voice-agent-evaluation | NVIDIA NeMo Voice Agent Evaluation | Reproducible live voice-agent harness for EVA and tau2 task domains | yes | yes | yes | yes | yes | no | no | yes | yes | Apache-2.0 | MIT for ported EVA and tau2 fixtures; Apache-2.0 for original project material | https://github.com/NVIDIA-NeMo/labs-Voice-Agent | null | null | 2026-09-01T00:00:00 | The repository ships a live bot-to-bot audio harness with 328 ported EVA and tau2 scenarios, tool execution, stateful outcome checks, resumable runs, and six documented success signals. It is a reproducible implementation layer over existing benchmark domains rather than a new independent task corpus. |
openbench | OpenBench | Reproducible ASR, diarization, orchestration, and streaming-transcription benchmarks | yes | no | no | no | no | no | yes | yes | yes | MIT | varies | https://github.com/argmaxinc/OpenBench | null | null | 2026-09-01T00:00:00 | OpenBench measures speech infrastructure and model quality, including streaming and diarization. It is not an end-to-end agent task-completion benchmark. |
openbenchmarks-voice-agent-latency | OpenBenchmarks Voice Agent Latency | Caller-perceived time to first agent audio measured from real phone calls | yes | yes | yes | no | no | no | no | yes | yes | MIT | CC-BY-4.0 | https://github.com/openbenchmarks-labs/voice-agent-latency | null | null | 2026-09-01T00:00:00 | The public mirror provides the caller harness, offline analyzer, per-turn artifacts, configuration receipts, recording references and checksums, and a verifier for TTFAB measured from real phone-call audio. It deliberately measures latency only, not answer quality, interruption handling, task completion, or feature bre... |
mu-bench | mu-bench | Multilingual customer-service ASR | yes | no | no | no | no | no | no | no | yes | Apache-2.0 | CC-BY-NC-4.0 | https://github.com/sierra-research/mu-bench | https://huggingface.co/datasets/sierra-research/mu-bench | null | 2026-09-01T00:00:00 | Public, gated audio covers 4,270 real customer-service utterances across five locales for ASR provider evaluation. The repository specifies Apache-2.0 for code and CC BY-NC 4.0 for data; this is not a spoken-agent action benchmark. |
elitr-bench | ELITR-Bench | Long-context LLM evaluation on meeting transcripts | no | no | yes | no | partial | no | yes | no | yes | BSD-3-Clause main code; Apache-2.0 notices for specified third-party files | CC-BY-4.0 | https://github.com/utter-project/ELITR-Bench | null | https://arxiv.org/abs/2403.20262 | 2026-09-01T00:00:00 | The benchmark evaluates long-context language-model behavior over meeting transcripts in single- and multi-turn modes. It starts from text transcripts rather than measuring audio capture, ASR, or computer action; the repository publishes separate code and data license files. |
audio-agent-bench-suite | Audio Agent Bench Suite | A collection of multi-turn spoken-agent benchmarks | yes | unclear | yes | yes | partial | no | no | unclear | partial | not stated | CC-BY-4.0 | null | https://huggingface.co/datasets/arcada-labs/audio-agent-bench-suite | null | 2026-09-01T00:00:00 | The public card links six spoken benchmark datasets spanning instruction following, knowledge grounding, function calls, memory, and state. The suite repository itself exposes the card but no substantive data files; its documented scoring compares responses with golden text and does not clearly establish spoken-output ... |
Voice Agent Benchmark Landscape
A structured, source-linked map of public benchmarks for voice agents, spoken assistants, speech-enabled tool use, computer action, ASR, and meeting understanding.
This is a landscape dataset, not a leaderboard. Each row records whether a benchmark covers spoken input/output, multi-turn interaction, tools, goal completion, computer or browser action, meeting or long-form content, real-time operation, and public data.
Values are deliberately conservative:
yes: explicitly supported or evaluated by the public source;no: outside the published scope or absent from the available evaluation;partial: present in only part of the suite or evaluated indirectly;unclear: public material was insufficient to classify confidently.
Facts were last verified on 2026-09-01. Evidence notes and first-party source links are included in every record.
The full methodology, contribution guide, and validator are available in the GitHub repository.
Versioned releases are permanently archived on Zenodo. The current archived release is v1.0.2.
Archived release date: 2026-09-01.
Load and filter the data
from datasets import load_dataset
landscape = load_dataset(
"chatjesus/voice-agent-benchmark-landscape",
split="train",
)
# Spoken benchmarks that explicitly evaluate tool use.
tool_use = landscape.filter(
lambda row: row["audio_input"] == "yes" and row["tool_use"] == "yes"
)
# Meeting or long-form starting points, preserving source URLs.
meeting = landscape.filter(lambda row: row["meeting_or_long_form"] == "yes")
print(meeting.select_columns(["name", "primary_focus", "github_url", "paper_url"]))
For shell workflows, the GitHub repository also includes a dependency-free query CLI with Markdown, JSON, JSONL, and CSV output.
Fields
id,name,primary_focus- capability fields:
audio_input,audio_output,multi_turn,tool_use,goal_completion,computer_or_browser_action,meeting_or_long_form,real_time,public_data code_license,data_licensegithub_url,huggingface_url,paper_urllast_verified,evidence_notes
How to use the landscape
Start with the user outcome rather than selecting a benchmark by name:
- Action-taking voice agents: combine task completion with tool accuracy, live interaction quality, recovery, confirmation, and side-effect safety. VoiceAgentBench, Audio2Tool, EVA, TalkAct, tau2-bench, and the NVIDIA NeMo evaluation harness cover complementary parts of this stack; OpenBenchmarks adds a separately reproducible caller-latency layer.
- Voice typing and dictation: combine ASR and streaming measures with semantic errors, entity accuracy, formatting, correction burden, and application insertion reliability. OpenBench and mu-bench are useful public starting points, while VoiceBench adds spoken instruction-following and reasoning coverage; none replaces product-specific desktop tests.
- Meeting transcription: measure capture, diarization, transcript accuracy, and grounded meeting understanding separately. OpenBench and ELITR-Bench address different layers of this problem.
The full benchmark selection guide and source audit explain the tradeoffs and evidence behind each row.
Maintainer
Maintained by Sophon LLC, makers of Cue — a desktop voice agent for voice typing, meeting transcription, and cross-app action. Free to start + Cue Plus $19.99/month.
No benchmark owner has sponsored inclusion.
- Downloads last month
- 116