aic_drive_thru_intercom_en

An evaluation dataset of drive-thru orders recorded through a low-cost intercom, for speech recognition and voice activity detection in the conditions a real ordering lane produces.

Each clip is a customer ordering from a car, captured by a cheap intercom microphone and preamplifier in a real noisy environment with engine and other noises. The dataset contains human ground-truth transcripts and hand-drawn VAD labels. No clean speech reference exists; evaluation is transcript- and VAD-based.

Dataset Description

Direct Use

Use it to evaluate speech enhancement, ASR and VAD systems on far-field, band-limited, noisy speech of the kind drive-thru and kiosk voice agents receive.

Technical Details

• 34 clips with a total duration of 00:14:32 [hh:mm:ss]
  • Minimum duration: 3.67 s
  • Maximum duration: 40.15 s
  • Mean duration: 25.65 s
• 48 kHz sampling rate, 16-bit, mono
• 4 speakers, all recorded with informed consent
• One speaker per clip: clips in which anyone other than the customer talks are excluded
• Voice gender distribution: 47 % female-sounding voices, 53 % male-sounding voices by clip count (3 of 4 speakers female-sounding)
• 1,281 transcript words

Dataset Structure

Each row in the dataset contains the following fields:

• mix: 48 kHz WAV audio of the intercom recording
• transcript: Ground-truth transcription of the customer's speech in mix
• id: Opaque unique sample identifier (random short uid)
• speaker_id: Numeric identifier of the customer (foreground speaker)
• index: Integer index distinguishing multiple samples from the same speaker
• menu_type: What is being ordered: burger, coffee, sandwich or taco
• mix__vad__human: Human voice-activity ground truth for the mix audio. It is an .npz marking where the customer is talking, at one label per 10 ms. See Human VAD labels

Dataset Sources

Audio

Recordings of a customer ordering through a low-cost drive-thru intercom chain:

• inexpensive intercom microphone with limited directivity
• intercom amplifier with high distortion. Output recorded via field recorder
• speakers were sitting in a car during the recording
• car-engine noises played back through a loudspeaker
• a busy street as the acoustic scene

Scenes vary across clips:

• customer far from the microphone
• customer turning away from or toward the microphone
• radio playing
• soft-spoken or very quiet customers

The result carries the pathologies of real intercom hardware: bandwidth limits, gain-staging problems, distortion and clipping, and a signal-to-noise ratio that changes within a clip. The orders are role-played. This is a controlled simulation, not live production traffic.

Transcriptions

• Human-annotated, and quality-checked by AIC staff.
• Each transcript covers the customer, the only talker in the clip.
• Menu items are transcribed as spoken.

VAD labels

• Annotated by hand by AIC staff against the mix audio.
• Aligned to the untrimmed mix audio, so they are row-aligned with every other field.

Dataset Details

• Curated by: Joschka Wohlgemuth
• Funded by: ai-coustics GmbH
• Language: English
• License: CC BY-NC 4.0
• Contact:
  • Email: info@ai-coustics.com
  • Web: https://ai-coustics.com

Limitations

• Small: 34 clips from 4 speakers; one speaker contributes about half the clips.
• No overlapping or competing talkers, so it does not test target-speaker separation.
• English only.
• Simulated ordering scenario recorded at one location with one hardware chain.

Citation

@dataset{aic_drive_thru_intercom_en,
  title        = {aic_drive_thru_intercom_en: An evaluation dataset of drive-thru intercom orders for speech recognition and voice activity detection},
  author       = {Joschka Wohlgemuth},
  year         = {2026},
  publisher    = {ai-coustics GmbH},
  license      = {CC BY-NC 4.0},
  url          = {https://ai-coustics.com}
}

Human VAD labels

mix__vad__human carries human voice-activity ground truth for every clip. Each cell is a {bytes, path} struct holding one .npz:

key	type	meaning
frame_rate_hz	float scalar, 100.0	VAD frames per second
foreground_speaker	bool array (n_frames,)	the ordering customer only

n_frames equals ceil(duration * 100) for every clip. The trailing partial 10 ms window carries a label, so the track can be one frame longer than the audio. The customer, the only talker in every clip, is active on 58.1 % of frames across the set.

import io, numpy as np
from datasets import load_dataset

ds = load_dataset("ai-coustics/aic_drive_thru_intercom_en", split="eval")
with np.load(io.BytesIO(ds[0]["mix__vad__human"]["bytes"]), allow_pickle=False) as d:
    fg = d["foreground_speaker"]  # bool, one entry per 10 ms
    fr = float(d["frame_rate_hz"])  # 100.0