The CMU-AIST submission for the ICME 2025 Audio Encoder Challenge

This technical report describes our submission to the ICME 2025 audio encoder challenge. Our submitted system is built on BEATs, a masked speech token prediction based audio encoder. We extend the BEATs model using 74,000 hours of data derived from various speech, music, and sound corpora and scale its architecture upto 300 million parameters. We experiment with speech-heavy and balanced pre-training mixtures to study the impact of different domains on final performance. Our submitted system consists of an ensemble of the Dasheng 1.2 billion model with two custom scaled-up BEATs models trained on the aforementioned pre-training data mixtures. We also propose a simple ensembling technique that retains the best capabilities of constituent models and surpasses both the baseline and Dasheng 1.2B. For open science, we publicly release our trained checkpoints via huggingface at https://huggingface.co/shikhar7ssu/OpenBEATs-ICME-SOUND and https://huggingface.co/shikhar7ssu/OpenBEATs-ICME.

Qwen3-TTS Technical Report

In this report, we present the Qwen3-TTS series, a family of advanced multilingual, controllable, robust, and streaming text-to-speech models. Qwen3-TTS supports state-of-the-art 3-second voice cloning and description-based control, allowing both the creation of entirely novel voices and fine-grained manipulation over the output speech. Trained on over 5 million hours of speech data spanning 10 languages, Qwen3-TTS adopts a dual-track LM architecture for real-time synthesis, coupled with two speech tokenizers: 1) Qwen-TTS-Tokenizer-25Hz is a single-codebook codec emphasizing semantic content, which offers seamlessly integration with Qwen-Audio and enables streaming waveform reconstruction via a block-wise DiT. 2) Qwen-TTS-Tokenizer-12Hz achieves extreme bitrate reduction and ultra-low-latency streaming, enabling immediate first-packet emission ($97\,\mathrm{ms}$) through its 12.5 Hz, 16-layer multi-codebook design and a lightweight causal ConvNet. Extensive experiments indicate state-of-the-art performance across diverse objective and subjective benchmark (e.g., TTS multilingual test set, InstructTTSEval, and our long speech test set). To facilitate community research and development, we release both tokenizers and models under the Apache 2.0 license.

Sink or SWIM: Tackling Real-Time ASR at Scale

Real-time automatic speech recognition systems are increasingly integrated into interactive applications, from voice assistants to live transcription services. However, scaling these systems to support multiple concurrent clients while maintaining low latency and high accuracy remains a major challenge. In this work, we present SWIM, a novel real-time ASR system built on top of OpenAI's Whisper model that enables true model-level parallelization for scalable, multilingual transcription. SWIM supports multiple concurrent audio streams without modifying the underlying model. It introduces a buffer merging strategy that maintains transcription fidelity while ensuring efficient resource usage. We evaluate SWIM in multi-client settings -- scaling up to 20 concurrent users -- and show that it delivers accurate real-time transcriptions in English, Italian, and Spanish, while maintaining low latency and high throughput. While Whisper-Streaming achieves a word error rate of approximately 8.2% with an average delay of approximately 3.4 s in a single-client, English-only setting, SWIM extends this capability to multilingual, multi-client environments. It maintains comparable accuracy with significantly lower delay -- around 2.4 s with 5 clients -- and continues to scale effectively up to 20 concurrent clients without degrading transcription quality and increasing overall throughput. Our approach advances scalable ASR by improving robustness and efficiency in dynamic, multi-user environments.

TidyVoice: A Curated Multilingual Dataset for Speaker Verification Derived from Common Voice

The development of robust, multilingual speaker recognition systems is hindered by a lack of large-scale, publicly available and multilingual datasets, particularly for the read-speech style crucial for applications like anti-spoofing. To address this gap, we introduce the TidyVoice dataset derived from the Mozilla Common Voice corpus after mitigating its inherent speaker heterogeneity within the provided client IDs. TidyVoice currently contains training and test data from over 212,000 monolingual speakers (Tidy-M) and around 4,500 multilingual speakers (Tidy-X) from which we derive two distinct conditions. The Tidy-M condition contains target and non-target trials from monolingual speakers across 81 languages. The Tidy-X condition contains target and non-target trials from multilingual speakers in both same- and cross-language trials. We employ two architectures of ResNet models, achieving a 0.35% EER by fine-tuning on our comprehensive Tidy-M partition. Moreover, we show that this fine-tuning enhances the model's generalization, improving performance on unseen conversational interview data from the CANDOR corpus. The complete dataset, evaluation trials, and our models are publicly released to provide a new resource for the community.

DeepASMR: LLM-Based Zero-Shot ASMR Speech Generation for Anyone of Any Voice

While modern Text-to-Speech (TTS) systems achieve high fidelity for read-style speech, they struggle to generate Autonomous Sensory Meridian Response (ASMR), a specialized, low-intensity speech style essential for relaxation. The inherent challenges include ASMR's subtle, often unvoiced characteristics and the demand for zero-shot speaker adaptation. In this paper, we introduce DeepASMR, the first framework designed for zero-shot ASMR generation. We demonstrate that a single short snippet of a speaker's ordinary, read-style speech is sufficient to synthesize high-fidelity ASMR in their voice, eliminating the need for whispered training data from the target speaker. Methodologically, we first identify that discrete speech tokens provide a soft factorization of ASMR style from speaker timbre. Leveraging this insight, we propose a two-stage pipeline incorporating a Large Language Model (LLM) for content-style encoding and a flow-matching acoustic decoder for timbre reconstruction. Furthermore, we contribute DeepASMR-DB, a comprehensive 670-hour English-Chinese multi-speaker ASMR speech corpus, and introduce a novel evaluation protocol integrating objective metrics, human listening tests, LLM-based scoring and unvoiced speech analysis. Extensive experiments confirm that DeepASMR achieves state-of-the-art naturalness and style fidelity in ASMR generation for anyone of any voice, while maintaining competitive performance on normal speech synthesis.

Adaptive Rotary Steering with Joint Autoregression for Robust Extraction of Closely Moving Speakers in Dynamic Scenarios

Latest advances in deep spatial filtering for Ambisonics demonstrate strong performance in stationary multi-speaker scenarios by rotating the sound field toward a target speaker prior to multi-channel enhancement. For applicability in dynamic acoustic conditions with moving speakers, we propose to automate this rotary steering using an interleaved tracking algorithm conditioned on the target's initial direction. However, for nearby or crossing speakers, robust tracking becomes difficult and spatial cues less effective for enhancement. By incorporating the processed recording as additional guide into both algorithms, our novel joint autoregressive framework leverages temporal-spectral correlations of speech to resolve spatially challenging speaker constellations. Consequently, our proposed method significantly improves tracking and enhancement of closely spaced speakers, consistently outperforming comparable non-autoregressive methods on a synthetic dataset. Real-world recordings complement these findings in complex scenarios with multiple speaker crossings and varying speaker-to-array distances.

Performance and Complexity Trade-off Optimization of Speech Models During Training

In speech machine learning, neural network models are typically designed by choosing an architecture with fixed layer sizes and structure. These models are then trained to maximize performance on metrics aligned with the task's objective. While the overall architecture is usually guided by prior knowledge of the task, the sizes of individual layers are often chosen heuristically. However, this approach does not guarantee an optimal trade-off between performance and computational complexity; consequently, post hoc methods such as weight quantization or model pruning are typically employed to reduce computational cost. This occurs because stochastic gradient descent (SGD) methods can only optimize differentiable functions, while factors influencing computational complexity, such as layer sizes and floating-point operations per second (FLOP/s), are non-differentiable and require modifying the model structure during training. We propose a reparameterization technique based on feature noise injection that enables joint optimization of performance and computational complexity during training using SGD-based methods. Unlike traditional pruning methods, our approach allows the model size to be dynamically optimized for a target performance-complexity trade-off, without relying on heuristic criteria to select which weights or structures to remove. We demonstrate the effectiveness of our method through three case studies, including a synthetic example and two practical real-world applications: voice activity detection and audio anti-spoofing. The code related to our work is publicly available to encourage further research.

Contrastive Knowledge Distillation for Embedding Refinement in Personalized Speech Enhancement

Personalized speech enhancement (PSE) has shown convincing results when it comes to extracting a known target voice among interfering ones. The corresponding systems usually incorporate a representation of the target voice within the enhancement system, which is extracted from an enrollment clip of the target voice with upstream models. Those models are generally heavy as the speaker embedding's quality directly affects PSE performances. Yet, embeddings generated beforehand cannot account for the variations of the target voice during inference time. In this paper, we propose to perform on-thefly refinement of the speaker embedding using a tiny speaker encoder. We first introduce a novel contrastive knowledge distillation methodology in order to train a 150k-parameter encoder from complex embeddings. We then use this encoder within the enhancement system during inference and show that the proposed method greatly improves PSE performances while maintaining a low computational load.

VCNAC: A Variable-Channel Neural Audio Codec for Mono, Stereo, and Surround Sound

We present VCNAC, a variable channel neural audio codec. Our approach features a single encoder and decoder parametrization that enables native inference for different channel setups, from mono speech to cinematic 5.1 channel surround audio. Channel compatibility objectives ensure that multi-channel content maintains perceptual quality when decoded to fewer channels. The shared representation enables training of generative language models on a single set of codebooks while supporting inference-time scalability across modalities and channel configurations. Evaluation using objective spatial audio metrics and subjective listening tests demonstrates that our unified approach maintains high reconstruction quality across mono, stereo, and surround audio configurations.

Deaf and Hard of Hearing Access to Intelligent Personal Assistants: Comparison of Voice-Based Options with an LLM-Powered Touch Interface

We investigate intelligent personal assistants (IPAs) accessibility for deaf and hard of hearing (DHH) people who can use their voice in everyday communication. The inability of IPAs to understand diverse accents including deaf speech renders them largely inaccessible to non-signing and speaking DHH individuals. Using an Echo Show, we compare the usability of natural language input via spoken English; with Alexa's automatic speech recognition and a Wizard-of-Oz setting with a trained facilitator re-speaking commands against that of a large language model (LLM)-assisted touch interface in a mixed-methods study. The touch method was navigated through an LLM-powered "task prompter," which integrated the user's history and smart environment to suggest contextually-appropriate commands. Quantitative results showed no significant differences across both spoken English conditions vs LLM-assisted touch. Qualitative results showed variability in opinions on the usability of each method. Ultimately, it will be necessary to have robust deaf-accented speech recognized natively by IPAs.