Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Treble10: A high-quality dataset for far-field speech recognition, dereverberation, and enhancement

Sarabeth S. Mullins, Georg Götz, Eric Bezzam, Steven Zheng, Daniel Gert Nielsen

TL;DR

Treble10 tackles the need for scalable, physically accurate far-field room-acoustic data to advance ASR, dereverberation, and enhancement. It introduces six subsets of RIRs and pre-convolved scenes across 10 furnished rooms, generated with a hybrid wave-based and geometrical-acoustics solver and sampled at $32 kHz$, with a transition between methods at $5 kHz$. The work includes LibriSpeech-based reverberant speech paired to each RIR format (mono, HOA8, 6-channel device) and makes the full dataset openly available on the Hugging Face Hub. The accompanying Treble SDK provides a programmable framework for reproducible, physics-grounded evaluation and large-scale data augmentation in far-field speech tasks.

Abstract

Accurate far-field speech datasets are critical for tasks such as automatic speech recognition (ASR), dereverberation, speech enhancement, and source separation. However, current datasets are limited by the trade-off between acoustic realism and scalability. Measured corpora provide faithful physics but are expensive, low-coverage, and rarely include paired clean and reverberant data. In contrast, most simulation-based datasets rely on simplified geometrical acoustics, thus failing to reproduce key physical phenomena like diffraction, scattering, and interference that govern sound propagation in complex environments. We introduce Treble10, a large-scale, physically accurate room-acoustic dataset. Treble10 contains over 3000 broadband room impulse responses (RIRs) simulated in 10 fully furnished real-world rooms, using a hybrid simulation paradigm implemented in the Treble SDK that combines a wave-based and geometrical acoustics solver. The dataset provides six complementary subsets, spanning mono, 8th-order Ambisonics, and 6-channel device RIRs, as well as pre-convolved reverberant speech scenes paired with LibriSpeech utterances. All signals are simulated at 32 kHz, accurately modelling low-frequency wave effects and high-frequency reflections. Treble10 bridges the realism gap between measurement and simulation, enabling reproducible, physically grounded evaluation and large-scale data augmentation for far-field speech tasks. The dataset is openly available via the Hugging Face Hub, and is intended as both a benchmark and a template for next-generation simulation-driven audio research.

Treble10: A high-quality dataset for far-field speech recognition, dereverberation, and enhancement

TL;DR

Treble10 tackles the need for scalable, physically accurate far-field room-acoustic data to advance ASR, dereverberation, and enhancement. It introduces six subsets of RIRs and pre-convolved scenes across 10 furnished rooms, generated with a hybrid wave-based and geometrical-acoustics solver and sampled at , with a transition between methods at . The work includes LibriSpeech-based reverberant speech paired to each RIR format (mono, HOA8, 6-channel device) and makes the full dataset openly available on the Hugging Face Hub. The accompanying Treble SDK provides a programmable framework for reproducible, physics-grounded evaluation and large-scale data augmentation in far-field speech tasks.

Abstract

Accurate far-field speech datasets are critical for tasks such as automatic speech recognition (ASR), dereverberation, speech enhancement, and source separation. However, current datasets are limited by the trade-off between acoustic realism and scalability. Measured corpora provide faithful physics but are expensive, low-coverage, and rarely include paired clean and reverberant data. In contrast, most simulation-based datasets rely on simplified geometrical acoustics, thus failing to reproduce key physical phenomena like diffraction, scattering, and interference that govern sound propagation in complex environments. We introduce Treble10, a large-scale, physically accurate room-acoustic dataset. Treble10 contains over 3000 broadband room impulse responses (RIRs) simulated in 10 fully furnished real-world rooms, using a hybrid simulation paradigm implemented in the Treble SDK that combines a wave-based and geometrical acoustics solver. The dataset provides six complementary subsets, spanning mono, 8th-order Ambisonics, and 6-channel device RIRs, as well as pre-convolved reverberant speech scenes paired with LibriSpeech utterances. All signals are simulated at 32 kHz, accurately modelling low-frequency wave effects and high-frequency reflections. Treble10 bridges the realism gap between measurement and simulation, enabling reproducible, physically grounded evaluation and large-scale data augmentation for far-field speech tasks. The dataset is openly available via the Hugging Face Hub, and is intended as both a benchmark and a template for next-generation simulation-driven audio research.
Paper Structure (8 sections, 3 figures, 1 table)

This paper contains 8 sections, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Motivation
  3. Near-field vs. far-field algorithm development
  4. Importance of domain data
  5. Importance of domain data
  6. How simulations can help
  7. The Treble10 dataset
  8. Treble Technologies and the Treble SDK

Figures (3)

  • Figure 1: The transformation of a clean into a reverberant audio signal via convolution with a simulated Room Impulse Response (RIR). The clean speech (left) is convolved with the RIR (centre). This operation yields augmented clean speech (right) that incorporates the acoustic characteristics of that particular room and source-receiver combination.
  • Figure 2: Disentangling the different degrees of freedom of room-acoustic datasets.
  • Figure 3: Sketch of the multi channel device included in the dataset. The device consists of $6$ microphones evenly spaced with a radius of 3cm. The coordinates of the microphones can be found in the metadata for the $6$ch split and also in the dataset card.