AHEAD-DS: Hearing-Aid Scene Recognition Benchmark

Updated 5 July 2026

AHEAD-DS is a fully specified, reproducible benchmark dataset for hearing-aid acoustic scene recognition featuring 9,968 10-second recordings across 14 audiologically relevant classes.
It addresses comparability issues by providing fixed train/validation/test splits, open-source mixing code, and standardized level normalization to balance diverse recording sources.
The accompanying YAMNet+ baseline model achieves a mAP of 0.83 and 0.93 accuracy, demonstrating effective transfer learning and feasible real-time performance on edge devices.

AHEAD-DS, short for Another HEaring AiD scenes Data Set, is a public, ready-to-use dataset for acoustic scene recognition in hearing-aid-relevant environments. It was introduced together with the lightweight baseline model YAMNet+ to provide a standardized, reproducible benchmark for scene recognition in hearing devices, addressing the lack of datasets that were simultaneously public, fully specified, already split into train/validation/test, and labeled in a way meaningful for hearing-device signal processing (Zhong et al., 14 Aug 2025). In its released form, AHEAD-DS contains 9968 recordings of 10 seconds, stored as single-channel, 16-bit integer, 16 kHz WAV files across 14 audiologically relevant classes; on its test set, YAMNet+ achieved mean average precision of 0.83 and accuracy of 0.93 (Zhong et al., 14 Aug 2025).

1. Origin and benchmark rationale

AHEAD-DS was designed to resolve a comparability problem in hearing-aid scene recognition research. Earlier datasets were described as often lacking public accessibility, completeness, reproducibility, or consistent audiologically meaningful labels. The immediate precursor was HEAR-DS, whose label set had already been motivated by audiological relevance and whose environmental recordings were open sourced under a permissive license. However, HEAR-DS was not released as a fixed benchmark: it was published in an unmixed state, involved pseudo-random procedures in dataset creation, and did not release the code needed to recover the exact mixed dataset or the exact train/test split (Zhong et al., 14 Aug 2025).

AHEAD-DS addresses those deficiencies by releasing the dataset in a ready-to-use final state, together with the mixing code and parameters, explicit ground-truth annotations, and predefined train/validation/test partitions. Its contribution is therefore not merely the curation of audio clips, but the establishment of a reproducible benchmark with audiologically targeted classes. The design objective is not generic acoustic-scene classification in the style of broad environmental taxonomies; rather, it is scene recognition for hearing devices, where class distinctions are intended to support different processing strategies such as filters, amplification, beamforming, noise suppression, and speech enhancement (Zhong et al., 14 Aug 2025).

The dataset was introduced in the paper "A dataset and model for recognition of audiologically relevant environments for hearing aids: AHEAD-DS and YAMNet+" (Zhong et al., 14 Aug 2025). In that formulation, the dataset and the model are paired: AHEAD-DS provides the benchmark, while YAMNet+ provides an open, edge-deployable baseline.

2. Label system and audiological semantics

AHEAD-DS adopts the same 14 labels as HEAR-DS. The authors retained this label set because it had already been selected as hearing-aid relevant by audiologists and a hearing-aid manufacturer, Audifon. The taxonomy is structured around environments and speech-in-environment conditions rather than around generic scene categories such as locations or object sounds (Zhong et al., 14 Aug 2025).

Type	Label
Pure environment	`cocktail_party`
Pure environment	`interfering_speakers`
Pure environment	`in_traffic`
Pure environment	`in_vehicle`
Pure environment	`music`
Pure environment	`quiet_indoors`
Pure environment	`reverberant_environment`
Pure environment	`wind_turbulence`
Speech in environment	`speech_in_traffic`
Speech in environment	`speech_in_vehicle`
Speech in environment	`speech_in_music`
Speech in environment	`speech_in_quiet_indoors`
Speech in environment	`speech_in_reverberant_env`
Speech in environment	`speech_in_wind_turbulence`

The rationale is explicitly audiological. The paper emphasizes that hearing devices must recognize environments in which different processing strategies may be appropriate, especially when speech is embedded in specific backgrounds. The class design therefore distinguishes, for example, speech_in_traffic, speech_in_vehicle, speech_in_music, and speech_in_reverberant_env, because speech intelligibility needs differ across those noise conditions (Zhong et al., 14 Aug 2025).

The dataset also encodes a specific distinction between two multi-speaker classes. interfering_speakers consists of speech-only material from CHiME 6 Dev, representing multiple speakers where speech remains intelligible, whereas cocktail_party represents multiple speakers but speech is unintelligible. This makes the taxonomy directly relevant to listening assistance rather than to conventional scene-label ontologies (Zhong et al., 14 Aug 2025).

The authors also identify a limitation in the label design. Although the labels are audiologically motivated, the field still lacks a more rigorous and systematic study of which hearing-relevant labels are optimally specific, too broad, or too narrow. This suggests that AHEAD-DS is both a benchmark and a provisional statement about task definition (Zhong et al., 14 Aug 2025).

3. Data sources, construction pipeline, and preprocessing

AHEAD-DS is assembled from two source datasets: HEAR-DS and CHiME 6 Dev. HEAR-DS provides the environmental sounds and music; CHiME 6 Dev provides the speech material (Zhong et al., 14 Aug 2025). The exact source summary given for the released benchmark is:

CHiME 6 Dev speech: 44 / 52 files used
HEAR-DS cocktail party: 1334 / 1334
HEAR-DS in traffic: 1056 / 1056
HEAR-DS in vehicle: 1168 / 1168
HEAR-DS music: 2992 / 2992
HEAR-DS quiet indoors: 1050 / 1050
HEAR-DS reverb env: 443 / 443
HEAR-DS wind turb: 878 / 878

The HEAR-DS environment recordings had been recorded with a dummy head using GRAS KB1065/1066 Pinnae, with an Audiofon microphone in each ear in an in-the-canal position, and a Focusrite Scarlett 18i6 for A/D conversion. They were initially captured at 48 kHz, 32-bit float, then downsampled to 16 kHz and quantized to 16-bit integer. From HEAR-DS, the paper reports 5929 clips of 10 s from 6 different environments, plus 2992 music clips of 10 s, from 10 genres originally sourced from GTZAN (Zhong et al., 14 Aug 2025).

Speech was drawn from CHiME 6 Dev rather than CHiME 5 because CHiME 6 Dev corrected recording synchronization while keeping the same speech content. The selected material comprised 44 sound files total, with 24 from one home environment and 20 from another, recorded during two dinner-party sessions in homes with open dining/kitchen/lounge spaces using Microsoft Kinects. Only Kinect audio was used; body-worn microphones were excluded because they sounded too clean and lacked acoustic realism. The recordings were 16-bit integer, 16 kHz (Zhong et al., 14 Aug 2025).

A major preprocessing issue was level mismatch across sources. CHiME 6 Dev and music were relatively high level, while other environmental recordings were much lower level. The authors therefore performed level standardization before mixing. During standardization and mixing, audio was converted to 32-bit floating point, then converted back to 16-bit integer for storage. The standardization strategy was:

for everything except CHiME 6 Dev and music: divide by the RMS of cocktail_party and multiply by the RMS of CHiME 6 Dev;
for music: divide by the RMS of music and multiply by the RMS of CHiME 6 Dev (Zhong et al., 14 Aug 2025).

The first second of each CHiME 6 Dev file was discarded because it contains a loud synchronization beep. The next 1100 seconds were then extracted in 10-second intervals, giving 110 clips per file, across 44 files, for a total of 4840 speech clips (Zhong et al., 14 Aug 2025).

The annotation logic is based on controlled source mixing. For each HEAR-DS environment used in speech-mixed categories, half of the clips were mixed with speech and half were left unmixed. Of the 4840 speech clips, 3793 were mixed with environmental sounds and 1047 were assigned directly to interfering_speakers. The classes cocktail_party and interfering_speakers were not mixed with other environment recordings. For every five speech/environment pairs, the authors created five mixes at $-10, -5, 0, 5, 10$ dB SNR, with speech as the signal and environment as the noise (Zhong et al., 14 Aug 2025).

An additional level-boosting step was introduced before applying the target SNR. If one component was quieter, it was boosted until it had equal RMS to the louder component, and speech was then adjusted to the desired SNR. This was introduced because some quiet_indoors clips had extremely low RMS; without that step, added speech at $-10$ dB SNR could become nearly all zeros after 16-bit quantization, creating speech-containing labels with imperceptible speech. The paper states that no speech clip was used more than once, reducing reuse leakage and preserving diversity among speech-containing classes (Zhong et al., 14 Aug 2025).

4. Benchmark structure, split policy, and reproducibility features

Every AHEAD-DS recording in the final release is 10 seconds, single-channel, 16-bit integer, and 16 kHz WAV, for a total of 9968 recordings (Zhong et al., 14 Aug 2025). After standardization and mixing, the dataset was split within each label into:

70% training
10% validation
20% testing

This produced a benchmark with 6980 training recordings, 1001 validation recordings, and 1987 testing recordings (Zhong et al., 14 Aug 2025).

AHEAD-DS is explicitly unbalanced. The largest classes are music and speech_in_music, each with 1496 clips, followed by cocktail_party with 1334 and interfering_speakers with 1047. The smallest classes are speech_in_reverberant_env with 221 and reverberant_environment with 222. This imbalance is one reason the accompanying baseline uses focal loss (Zhong et al., 14 Aug 2025).

For training, three augmentations were applied on the fly, each independently with probability 0.5:

Gain augmentation between $-6$ and $+6$ dB
Additive noise perturbation uniform in $[-0.003, 0.003]$
Time stretch/shrink by a factor between $0.9$ and $1.1$, using a Fourier resampler (Zhong et al., 14 Aug 2025)

The reproducibility features are central to the benchmark’s identity. The paper highlights that the mixing code is published, the mixing parameters are documented, the final dataset is released in fixed form, and both mixed and unmixed versions are available. Labels are stored in a spreadsheet, audio is stored as WAV, and the dataset is already partitioned into training/validation/testing. The recommended evaluation protocol is correspondingly conservative: use the predefined split, keep the released dataset fixed for comparability, report at least mAP and accuracy, and preserve the documented level-normalization assumptions (Zhong et al., 14 Aug 2025).

The release locations include a project website, a GitHub repository for YAMNet+, Hugging Face repositories for the mixed and unmixed datasets, and a model repository. The dataset and code are described as released under a permissive licence (Zhong et al., 14 Aug 2025).

5. YAMNet+ baseline model and reported benchmark performance

YAMNet+ is the baseline model introduced alongside AHEAD-DS. The authors considered PANN, AST, and YAMNet as candidate pretrained sound-recognition models. Although PANN and AST have better AudioSet mAP than YAMNet, they are much larger: YAMNet / YAMNet+ have about 3.7M parameters, whereas PANN has about 81M and AST about 86M. Because hearing-aid-adjacent applications may require deployment on resource-constrained edge devices, the smaller YAMNet architecture was chosen (Zhong et al., 14 Aug 2025).

YAMNet+ keeps the same internal architecture as YAMNet but adds a complete open-source workflow for loading weights, transfer learning, training, testing, and edge conversion. It uses raw waveform input, computes a log mel spectrogram internally, and employs a MobileNet backbone. The architecture consists of an initial convolutional block followed by 13 blocks containing depthwise separable convolution, batch normalization, ReLU, convolution, batch normalization, and ReLU, then a global average pooling penultimate layer and a final fully connected layer with sigmoid activation of length equal to the number of classes. The reported parameter counts are 3,751,369 when initialized with 521 classes and 3,743,831 when initialized with the 14-class output used for AHEAD-DS (Zhong et al., 14 Aug 2025).

The input waveforms are sliced into windows of 960 ms with 480 ms overlap, so a 10 s clip becomes 20 windows. If necessary, zero padding is applied. Each 960 ms window is classified separately, and the paper states that scores were not combined across windows during recognition; each window is treated as an independent instance (Zhong et al., 14 Aug 2025).

A major empirical conclusion is that transfer learning from the pretrained YAMNet model was essential. The best overall configuration used AudioSet-pretrained initialization with all layers trainable. Random initialization performed poorly and was interpreted as evidence that AHEAD-DS alone is too small to train YAMNet+ effectively from scratch. Training used a maximum of 100 epochs, the Adam optimizer, learning rate 0.00001, a decay-on-plateau scheduler, focal loss with $\alpha = 0.25$ and $\gamma = 2$ , and label smoothing with smoothing factor 0.1 (Zhong et al., 14 Aug 2025).

On the test set of AHEAD-DS, YAMNet+ achieved:

mAP = 0.83
accuracy = 0.93 (Zhong et al., 14 Aug 2025)

The confusion matrix analysis reported three prominent confusions:

speech_in_music vs music
speech_in_traffic vs in_traffic
speech_in_vehicle vs speech_in_wind_turbulence (Zhong et al., 14 Aug 2025)

These errors were interpreted in relation to the windowed recognition setup. Several paired classes differ only by the presence of speech, so 960 ms windows that fall into pauses naturally become ambiguous. The paper also reports a strong gain sensitivity effect: validation performance was best at 0 dB gain, ანუ at the standardized level expected by the model, and degraded away from that point. This reinforces the importance of source-level normalization when transferring the system to new recording hardware or datasets (Zhong et al., 14 Aug 2025).

The edge-deployment results are a notable part of the benchmark’s practical significance. After conversion to TensorFlow Lite 16-bit floating point, the model size was 6.19 MiB. On a Google Pixel 3 running Android 13, using a Flutter/Dart frontend with TensorFlow Lite inference, the mean processing time for 5 s of audio was 200.4 ms. The fitted linear latency model was:

$y = 29.62x + 48.31$

where $-10$ 0 is seconds of audio and $-10$ 1 is processing time in milliseconds. The paper interprets this as approximately 50 ms one-off latency to load the model and an approximate linear increase of 30 ms per 1 second of audio, concluding that real-time sound-based scene recognition is feasible on modest edge hardware (Zhong et al., 14 Aug 2025).

6. Limitations, future directions, and nomenclatural scope

The authors identify several limitations. First, AHEAD-DS is still small relative to datasets such as AudioSet, which helps explain the dependence on transfer learning, the underfitting observed under random initialization, and the limited gains from regularization. Second, the label system, while audiologically motivated, is not presented as a definitive taxonomy; the field still lacks a rigorous framework for deciding which scene classes are optimally relevant for hearing research and ecological momentary assessment. Third, the dataset inherits source and recording bias from a small number of corpora and recording setups, especially HEAR-DS environment recordings and CHiME 6 Dev domestic dinner-party speech. Fourth, the baseline model is sensitive to sound-level mismatch. Fifth, the transfer-learning setup effectively locks the system to 960 ms windows, which may not be ideal for all scene types. Sixth, the implementation depends on TensorFlow and legacy compatibility workarounds (Zhong et al., 14 Aug 2025).

The proposed future directions are correspondingly pragmatic: expand AHEAD-DS with more clips, more varied environments, and more recording devices and configurations; develop systematic label design criteria; investigate on-the-fly mixing for training while keeping the test set fixed; explore alternative architectures such as EfficientNet, ResNet, and possibly time-domain features for reverberation-sensitive tasks; move to more modern ML frameworks such as JAX or PyTorch; and test inference on a broader range of edge devices, potentially moving toward deployment on hearing devices themselves (Zhong et al., 14 Aug 2025).

A further issue is nomenclatural. Within the literature considered here, AHEAD-DS explicitly denotes the hearing-aid acoustic-scene dataset Another HEaring AiD scenes Data Set (Zhong et al., 14 Aug 2025). Other arXiv works with superficially similar names are not formal definitions of AHEAD-DS. "Anticipation-driven Adaptive Architecture for Assisted Living" describes ADAA and is relevant only as an architectural analogue in assisted living; it does not explicitly mention the name “AHEAD-DS” (Nadin et al., 2021). "Adaptive Draft-Verification for Efficient LLM Decoding" introduces ADED, not AHEAD-DS (Liu et al., 2024). A plausible implication is that the dataset paper establishes the explicit and stable usage of the term in the hearing-aid acoustic-scene-recognition context, while similar acronyms in other domains should be treated as distinct entities.

In that sense, AHEAD-DS is best understood as a reproducible benchmark for hearing-aid-relevant acoustic scene recognition: a public, fixed, labeled, and split dataset coupled to an open, lightweight baseline that supports both research comparability and practical edge deployment (Zhong et al., 14 Aug 2025).