We present BreezyVoice, a Text-to-Speech (TTS) system specifically adapted for Taiwanese Mandarin, highlighting phonetic control abilities to address the unique challenges of polyphone disambiguation in the language. Building upon CosyVoice, we incorporate a $S^{3}$ tokenizer, a large language model (LLM), an optimal-transport conditional flow matching model (OT-CFM), and a grapheme to phoneme prediction model, to generate realistic speech that closely mimics human utterances. Our evaluation demonstrates BreezyVoice's superior performance in both general and code-switching contexts, highlighting its robustness and effectiveness in generating high-fidelity speech. Additionally, we address the challenges of generalizability in modeling long-tail speakers and polyphone disambiguation. Our approach significantly enhances performance and offers valuable insights into the workings of neural codec TTS systems.

This technical report presents our initial attempt to build a spoken large language model (LLM) for Taiwanese Mandarin, specifically tailored to enable real-time, speech-to-speech interaction in multi-turn conversations. Our end-to-end model incorporates a decoder-only transformer architecture and aims to achieve seamless interaction while preserving the conversational flow, including full-duplex capabilities allowing simultaneous speaking and listening. The paper also details the training process, including data preparation with synthesized dialogues and adjustments for real-time interaction. We also developed a platform to evaluate conversational fluency and response coherence in multi-turn dialogues. We hope the release of the report can contribute to the future development of spoken LLMs in Taiwanese Mandarin.

Multimodal foundation models, such as Gemini and ChatGPT, have revolutionized human-machine interactions by seamlessly integrating various forms of data. Developing a universal spoken language model that comprehends a wide range of natural language instructions is critical for bridging communication gaps and facilitating more intuitive interactions. However, the absence of a comprehensive evaluation benchmark poses a significant challenge. We present Dynamic-SUPERB Phase-2, an open and evolving benchmark for the comprehensive evaluation of instruction-based universal speech models. Building upon the first generation, this second version incorporates 125 new tasks contributed collaboratively by the global research community, expanding the benchmark to a total of 180 tasks, making it the largest benchmark for speech and audio evaluation. While the first generation of Dynamic-SUPERB was limited to classification tasks, Dynamic-SUPERB Phase-2 broadens its evaluation capabilities by introducing a wide array of novel and diverse tasks, including regression and sequence generation, across speech, music, and environmental audio. Evaluation results indicate that none of the models performed well universally. SALMONN-13B excelled in English ASR, while WavLLM demonstrated high accuracy in emotion recognition, but current models still require further innovations to handle a broader range of tasks. We will soon open-source all task data and the evaluation pipeline.

Despite their impressive success, training foundation models remains computationally costly. This paper investigates how to efficiently train speech foundation models with self-supervised learning (SSL) under a limited compute budget. We examine critical factors in SSL that impact the budget, including model architecture, model size, and data size. Our goal is to make analytical steps toward understanding the training dynamics of speech foundation models. We benchmark SSL objectives in an entirely comparable setting and find that other factors contribute more significantly to the success of SSL. Our results show that slimmer model architectures outperform common small architectures under the same compute and parameter budget. We demonstrate that the size of the pre-training data remains crucial, even with data augmentation during SSL training, as performance suffers when iterating over limited data. Finally, we identify a trade-off between model size and data size, highlighting an optimal model size for a given compute budget.

Speech Integrated Large Language Models (SILLMs) combine large language models with speech perception to perform diverse tasks, such as emotion recognition to speaker verification, demonstrating universal audio understanding capability. However, these models may amplify biases present in training data, potentially leading to biased access to information for marginalized groups. This work introduces a curated spoken bias evaluation toolkit and corresponding dataset. We evaluate gender bias in SILLMs across four semantic-related tasks: speech-to-text translation (STT), spoken coreference resolution (SCR), spoken sentence continuation (SSC), and spoken question answering (SQA). Our analysis reveals that bias levels are language-dependent and vary with different evaluation methods. Our findings emphasize the necessity of employing multiple approaches to comprehensively assess biases in SILLMs, providing insights for developing fairer SILLM systems.

Self-supervised learning (SSL) speech models have achieved remarkable performance in various tasks, yet the biased outcomes, especially affecting marginalized groups, raise significant concerns. Social bias refers to the phenomenon where algorithms potentially amplify disparate properties between social groups present in the data used for training. Bias in SSL models can perpetuate injustice by automating discriminatory patterns and reinforcing inequitable systems. This work reveals that prevalent SSL models inadvertently acquire biased associations. We probe how various factors, such as model architecture, size, and training methodologies, influence the propagation of social bias within these models. Finally, we explore the efficacy of debiasing SSL models through regularization techniques, specifically via model compression. Our findings reveal that employing techniques such as row-pruning and training wider, shallower models can effectively mitigate social bias within SSL model.