Datasets:

fluid-concepts
/

tooltalk-samples

pretty_name: >-
  ToolTalk Samples - High Quality Duplex Speech and Tool-calling in Customer
  Service Domain
license: other
license_name: fluid-concepts-sample-terms
license_link: https://huggingface.co/datasets/fluid-concepts/tooltalk-samples#terms
language:
  - en
tags:
  - audio
  - speech
  - dialogue
  - tool-use
  - function-calling
  - customer-service
  - tooltalk
  - airline
  - food_delivery
  - healthcare_appts
  - hotel
  - restaurant
  - roadside_assistance
  - takeout
size_categories:
  - n<1K
configs:
  - config_name: clean
    default: true
    data_files:
      - split: train
        path: sets/clean/data/**
  - config_name: noisy
    data_files:
      - split: train
        path: sets/noisy/data/**
extra_gated_heading: Access to these samples
extra_gated_prompt: >-
  These recordings are samples shared under the terms set out in the Terms
  section of this page: evaluation only, no redistribution, no cloning of
  anyone's voice or likeness. Tell us who you are and how you intend to use
  them; access is granted at once to the account that agrees.
extra_gated_fields:
  Name: text
  Company: text
  Intended use: text
  I agree to the terms of use on this page: checkbox
extra_gated_button_content: Agree and access

ToolTalk Samples - High Quality Duplex Speech and Tool-calling in Customer Service Domain

Two people improvise realistic customer-service calls while one operates a live, stateful tool environment—with synchronized speaker-separated audio, tool calls, and outcomes.

▶ Listen to Clean · ▶ Listen to Noisy · Discuss the full dataset

In this sample: 26 calls · 78.6 minutes · 7 sample domains · 203 tool calls

Technical specs: 48 kHz / 32-bit PCM speaker-separated source tracks · sub −60 dB recording noise floor · RT60 < 90 ms · no compression or EQ · raw audio · training-ready processing on request

What makes ToolTalk different?

Human interaction, not generated dialogue. Two real people improvise each call from a persona and objective; conversations are not scripted line by line.
Live, stateful tools. The agent works against domain databases and functions that read and change state, producing checkable operational outcomes.
Everything shares one clock. Speaker-separated audio, overlaps, tool arguments and results, and agent behavior are synchronized to the same timeline.
Noise changes the conversation, not just the waveform. In Noisy calls, participants hear the aligned street or freeway noise bed while speaking and operating the tools; clean microphone tracks and the noise bed remain separately available.
Raw by default, training-ready on request. This is fully raw audio, which gives you the most flexibility to preprocess for your own pipeline. We can also provide training-ready audio: normalized, equalized and denoised for high-fidelity speech generation.

Clean and Noisy sets

Set	Calls	Minutes	Tool calls	Included	Listen
Clean	14	43.8	109	Separate agent and customer studio tracks plus a clean combined stereo call	Open Clean viewer
Noisy	12	34.8	94	Separate clean speech tracks, the aligned ambient bed, and a post-processed combined call	Open Noisy viewer

Sample domains: ✈️ Airline reservations · 🛵 Food delivery · ❤️‍🩹 Healthcare appointment booking · 🏨 Hotel booking · 🪑 Restaurant reservations · 🍝 Restaurant take-out and pick-up · 🚗 Roadside assistance

The full ToolTalk collection also includes banking.

Technical documentation

See TECHNICAL.md for recording specifications, file layouts, noise processing, transcript formats, the complete schema, tool/audio synchronization, and loading examples.

Terms

Fluid Concepts provides these recordings solely for evaluation by people who have agreed to these terms on the data page.

Evaluation only. Use the samples to evaluate the ToolTalk collection for a possible engagement with Fluid Concepts. Do not use them to train, fine-tune or evaluate a model for any other purpose without written agreement.
No redistribution. Do not share the files, or derivatives of them, with anyone who has not agreed to these terms on the data page, and do not post them publicly.
People. The voices belong to paid participants recorded with their consent. Participant ids are anonymous studio codes; do not attempt to identify anyone.
No cloning. Do not use the recordings to clone, synthesize or imitate anyone's voice, or to build a digital likeness or avatar of any participant.
Revocation. Access can be withdrawn at any time. On request, delete all copies.
No warranty. The samples are provided as they are.

Questions: hello@fluidconcepts.ai

The same text is TERMS.md in this repository (public copy); requesting access is accepting it.