Po‐Chun Hsu

Paden Tomasello, Akshat Shrivastava, Daniel Lazar et al. · meta-ai, uw

End-to-end spoken language understanding (SLU) predicts intent directly from audio using a single model. It promises to improve the performance of assistant systems by leveraging acoustic information lost in the intermediate textual representation and preventing cascading errors from Automatic Speech Recognition (ASR). Further, having one unified model has efficiency advantages when deploying assistant systems on-device. However, the limited number of public audio datasets with semantic parse labels hinders the research progress in this area. In this paper, we release the Spoken Task-Oriented semantic Parsing (STOP) dataset, the largest and most complex SLU dataset to be publicly available. Additionally, we define low-resource splits to establish a benchmark for improving SLU when limited labeled data is available. Furthermore, in addition to the human-recorded audio, we are releasing a TTS-generated version to benchmark the performance for low-resource domain adaptation of end-to-end SLU systems. Initial experimentation show end-to-end SLU models performing slightly worse than their cascaded counterparts, which we hope encourages future work in this direction.

Chan-Jan Hsu, Chang-Le Liu, Feng-Ting Liao et al.

The evaluation of large language models is an essential task in the field of language understanding and generation. As language models continue to advance, the need for effective benchmarks to assess their performance has become imperative. In the context of Traditional Chinese, there is a scarcity of comprehensive and diverse benchmarks to evaluate the capabilities of language models, despite the existence of certain benchmarks such as DRCD, TTQA, CMDQA, and FGC dataset. To address this gap, we propose a novel set of benchmarks that leverage existing English datasets and are tailored to evaluate language models in Traditional Chinese. These benchmarks encompass a wide range of tasks, including contextual question-answering, summarization, classification, and table understanding. The proposed benchmarks offer a comprehensive evaluation framework, enabling the assessment of language models' capabilities across different tasks. In this paper, we evaluate the performance of GPT-3.5, Taiwan-LLaMa-v1.0, and Model 7-C, our proprietary model, on these benchmarks. The evaluation results highlight that our model, Model 7-C, achieves performance comparable to GPT-3.5 with respect to a part of the evaluated capabilities. In an effort to advance the evaluation of language models in Traditional Chinese and stimulate further research in this field, we have open-sourced our benchmark and opened the model for trial.

Philipp Ennen, Po-Chun Hsu, Chan-Jan Hsu et al.

In this paper we present the multilingual language model BLOOM-zh that features enhanced support for Traditional Chinese. BLOOM-zh has its origins in the open-source BLOOM models presented by BigScience in 2022. Starting from released models, we extended the pre-training of BLOOM by additional 7.4 billion tokens in Traditional Chinese and English covering a variety of domains such as news articles, books, encyclopedias, educational materials as well as spoken language. In order to show the properties of BLOOM-zh, both existing and newly created benchmark scenarios are used for evaluating the performance. BLOOM-zh outperforms its predecessor on most Traditional Chinese benchmarks while maintaining its English capability. We release all our models to the research community.

Chi-Luen Feng, Po-chun Hsu, Hung-yi Lee

Self-Supervised Learning (SSL) has made great strides recently. SSL speech models achieve decent performance on a wide range of downstream tasks, suggesting that they extract different aspects of information from speech. However, how SSL models store various information in hidden representations without interfering is still poorly understood. Taking the recently successful SSL model, HuBERT, as an example, we explore how the SSL model processes and stores speaker information in the representation. We found that HuBERT stores speaker information in representations whose positions correspond to silences in a waveform. There are several pieces of evidence. (1) We find that the utterances with more silent parts in the waveforms have better Speaker Identification (SID) accuracy. (2) If we use the whole utterances for SID, the silence part always contributes more to the SID task. (3) If we only use the representation of a part of the utterance for SID, the silenced part has higher accuracy than the other parts. Our findings not only contribute to a better understanding of SSL models but also improve performance. By simply adding silence to the original waveform, HuBERT improved its accuracy on SID by nearly 2%.

Da-rong Liu, Po-chun Hsu, Yi-chen Chen et al.

ASR has been shown to achieve great performance recently. However, most of them rely on massive paired data, which is not feasible for low-resource languages worldwide. This paper investigates how to learn directly from unpaired phone sequences and speech utterances. We design a two-stage iterative framework. GAN training is adopted in the first stage to find the mapping relationship between unpaired speech and phone sequence. In the second stage, another HMM model is introduced to train from the generator's output, which boosts the performance and provides a better segmentation for the next iteration. In the experiment, we first investigate different choices of model designs. Then we compare the framework to different types of baselines: (i) supervised methods (ii) acoustic unit discovery based methods (iii) methods learning from unpaired data. Our framework performs consistently better than all acoustic unit discovery methods and previous methods learning from unpaired data based on the TIMIT dataset.

Po-Chun Hsu, Li-Hsiang Shen, Chun-Hung Liu et al.

Terahertz (THz) communication with ultra-wide available spectrum is a promising technique that can achieve the stringent requirement of high data rate in the next-generation wireless networks, yet its severe propagation attenuation significantly hinders its implementation in practice. Finding beam directions for a large-scale antenna array to effectively overcome severe propagation attenuation of THz signals is a pressing need. This paper proposes a novel approach of federated deep reinforcement learning (FDRL) to swiftly perform THz-beam search for multiple base stations (BSs) coordinated by an edge server in a cellular network. All the BSs conduct deep deterministic policy gradient (DDPG)-based DRL to obtain THz beamforming policy with limited channel state information (CSI). They update their DDPG models with hidden information in order to mitigate inter-cell interference. We demonstrate that the cell network can achieve higher throughput as more THz CSI and hidden neurons of DDPG are adopted. We also show that FDRL with partial model update is able to nearly achieve the same performance of FDRL with full model update, which indicates an effective means to reduce communication load between the edge server and the BSs by partial model uploading. Moreover, the proposed FDRL outperforms conventional non-learning-based and existing non-FDRL benchmark optimization methods.

MediaTek Research, Chan-Jan Hsu, Chia-Sheng Liu et al.

Llama-Breeze2 (hereinafter referred to as Breeze2) is a suite of advanced multi-modal language models, available in 3B and 8B parameter configurations, specifically designed to enhance Traditional Chinese language representation. Building upon the Llama 3.2 model family, we continue the pre-training of Breeze2 on an extensive corpus to enhance the linguistic and cultural heritage of Traditional Chinese. In addition to language modeling capabilities, we significantly augment the models with function calling and vision understanding capabilities. At the time of this publication, as far as we are aware, absent reasoning-inducing prompts, Breeze2 are the strongest performing models in Traditional Chinese function calling and image understanding in its size class. The effectiveness of Breeze2 is benchmarked across various tasks, including Taiwan general knowledge, instruction-following, long context, function calling, and vision understanding. We are publicly releasing all Breeze2 models under the Llama 3.2 Community License. We also showcase the capabilities of the model running on mobile platform with a mobile application which we also open source.

Haibin Wu, Po-chun Hsu, Ji Gao et al.

Automatic speaker verification (ASV), one of the most important technology for biometric identification, has been widely adopted in security-critical applications. However, ASV is seriously vulnerable to recently emerged adversarial attacks, yet effective countermeasures against them are limited. In this paper, we adopt neural vocoders to spot adversarial samples for ASV. We use the neural vocoder to re-synthesize audio and find that the difference between the ASV scores for the original and re-synthesized audio is a good indicator for discrimination between genuine and adversarial samples. This effort is, to the best of our knowledge, among the first to pursue such a technical direction for detecting time-domain adversarial samples for ASV, and hence there is a lack of established baselines for comparison. Consequently, we implement the Griffin-Lim algorithm as the detection baseline. The proposed approach achieves effective detection performance that outperforms the baselines in all the settings. We also show that the neural vocoder adopted in the detection framework is dataset-independent. Our codes will be made open-source for future works to do fair comparison.

Chih-Kai Yang, Yu-Kuan Fu, Chen-An Li et al.

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.

Yi-Chang Chen, Po-Chun Hsu, Chan-Jan Hsu et al.

Large language models (LLMs) have significantly advanced autonomous agents, particularly in zero-shot tool usage, also known as function calling. This research delves into enhancing the function-calling capabilities of LLMs by exploring different approaches, including prompt formats for integrating function descriptions, blending function-calling and instruction-following data, introducing a novel Decision Token for conditional prompts, leveraging chain-of-thought reasoning, and overcoming multilingual challenges with a translation pipeline. Our key findings and contributions are as follows: (1) Instruction-following data improves both function-calling accuracy and relevance detection. (2) The use of the newly proposed Decision Token, combined with synthetic non-function-call data, enhances relevance detection. (3) A tailored translation pipeline effectively overcomes multilingual limitations, demonstrating significant improvements in Traditional Chinese. These insights highlight the potential for improved function-calling capabilities and multilingual applications in LLMs.

Chan-Jan Hsu, Yi-Chang Chen, Feng-Ting Liao et al.

We propose "Generative Fusion Decoding" (GFD), a novel shallow fusion framework designed to integrate large language models (LLMs) into cross-modal text recognition systems for automatic speech recognition (ASR) and optical character recognition (OCR). We derive the necessary formulations to enable GFD to operate across mismatched token spaces of different models by calculating likelihood at the byte level, thereby enabling seamless fusion and synchronous progression during the decoding process. GFD is plug-and-play by design, making it readily compatible with various auto-regressive models without the need for any re-training. GFD proves effective for general ASR and OCR tasks through intermediate and frequent interactions with LLMs, surpassing cascaded methods in English and Mandarin benchmarks. In addition, GFD transfers in-context learning abilities of LLMs and allows for adaptive ASR in instruction-aware and long-context settings, yielding significant WER reductions of up to 17.7\%.

Chung-Ming Chien, Jheng-Hao Lin, Chien-yu Huang et al.

The few-shot multi-speaker multi-style voice cloning task is to synthesize utterances with voice and speaking style similar to a reference speaker given only a few reference samples. In this work, we investigate different speaker representations and proposed to integrate pretrained and learnable speaker representations. Among different types of embeddings, the embedding pretrained by voice conversion achieves the best performance. The FastSpeech 2 model combined with both pretrained and learnable speaker representations shows great generalization ability on few-shot speakers and achieved 2nd place in the one-shot track of the ICASSP 2021 M2VoC challenge.

Po-chun Hsu, Hung-yi Lee

In this paper, we propose WG-WaveNet, a fast, lightweight, and high-quality waveform generation model. WG-WaveNet is composed of a compact flow-based model and a post-filter. The two components are jointly trained by maximizing the likelihood of the training data and optimizing loss functions on the frequency domains. As we design a flow-based model that is heavily compressed, the proposed model requires much less computational resources compared to other waveform generation models during both training and inference time; even though the model is highly compressed, the post-filter maintains the quality of generated waveform. Our PyTorch implementation can be trained using less than 8 GB GPU memory and generates audio samples at a rate of more than 960 kHz on an NVIDIA 1080Ti GPU. Furthermore, even if synthesizing on a CPU, we show that the proposed method is capable of generating 44.1 kHz speech waveform 1.2 times faster than real-time. Experiments also show that the quality of generated audio is comparable to those of other methods. Audio samples are publicly available online.

Po-chun Hsu, Chun-hsuan Wang, Andy T. Liu et al.

Recently, neural vocoders have been widely used in speech synthesis tasks, including text-to-speech and voice conversion. However, when encountering data distribution mismatch between training and inference, neural vocoders trained on real data often degrade in voice quality for unseen scenarios. In this paper, we train four common neural vocoders, including WaveNet, WaveRNN, FFTNet, Parallel WaveGAN alternately on five different datasets. To study the robustness of neural vocoders, we evaluate the models using acoustic features from seen/unseen speakers, seen/unseen languages, a text-to-speech model, and a voice conversion model. We found out that the speaker variety is much more important for achieving a universal vocoder than the language. Through our experiments, we show that WaveNet and WaveRNN are more suitable for text-to-speech models, while Parallel WaveGAN is more suitable for voice conversion applications. Great amount of subjective MOS results in naturalness for all vocoders are presented for future studies.

Andy T. Liu, Shu-wen Yang, Po-Han Chi et al.

We present Mockingjay as a new speech representation learning approach, where bidirectional Transformer encoders are pre-trained on a large amount of unlabeled speech. Previous speech representation methods learn through conditioning on past frames and predicting information about future frames. Whereas Mockingjay is designed to predict the current frame through jointly conditioning on both past and future contexts. The Mockingjay representation improves performance for a wide range of downstream tasks, including phoneme classification, speaker recognition, and sentiment classification on spoken content, while outperforming other approaches. Mockingjay is empirically powerful and can be fine-tuned with downstream models, with only 2 epochs we further improve performance dramatically. In a low resource setting with only 0.1% of labeled data, we outperform the result of Mel-features that uses all 100% labeled data.

Andy T. Liu, Po-chun Hsu, Hung-yi Lee

We present an unsupervised end-to-end training scheme where we discover discrete subword units from speech without using any labels. The discrete subword units are learned under an ASR-TTS autoencoder reconstruction setting, where an ASR-Encoder is trained to discover a set of common linguistic units given a variety of speakers, and a TTS-Decoder trained to project the discovered units back to the designated speech. We propose a discrete encoding method, Multilabel-Binary Vectors (MBV), to make the ASR-TTS autoencoder differentiable. We found that the proposed encoding method offers automatic extraction of speech content from speaker style, and is sufficient to cover full linguistic content in a given language. Therefore, the TTS-Decoder can synthesize speech with the same content as the input of ASR-Encoder but with different speaker characteristics, which achieves voice conversion (VC). We further improve the quality of VC using adversarial training, where we train a TTS-Patcher that augments the output of TTS-Decoder. Objective and subjective evaluations show that the proposed approach offers strong VC results as it eliminates speaker identity while preserving content within speech. In the ZeroSpeech 2019 Challenge, we achieved outstanding performance in terms of low bitrate.

Cheng-chieh Yeh, Po-chun Hsu, Ju-chieh Chou et al.

Speaking rate refers to the average number of phonemes within some unit time, while the rhythmic patterns refer to duration distributions for realizations of different phonemes within different phonetic structures. Both are key components of prosody in speech, which is different for different speakers. Models like cycle-consistent adversarial network (Cycle-GAN) and variational auto-encoder (VAE) have been successfully applied to voice conversion tasks without parallel data. However, due to the neural network architectures and feature vectors chosen for these approaches, the length of the predicted utterance has to be fixed to that of the input utterance, which limits the flexibility in mimicking the speaking rates and rhythmic patterns for the target speaker. On the other hand, sequence-to-sequence learning model was used to remove the above length constraint, but parallel training data are needed. In this paper, we propose an approach utilizing sequence-to-sequence model trained with unsupervised Cycle-GAN to perform the transformation between the phoneme posteriorgram sequences for different speakers. In this way, the length constraint mentioned above is removed to offer rhythm-flexible voice conversion without requiring parallel data. Preliminary evaluation on two datasets showed very encouraging results.