Recent advancements in Neural Audio Codec (NAC) models have inspired their use in various speech processing tasks, including speech enhancement (SE). In this work, we propose a novel, efficient SE approach by leveraging the pre-quantization output of a pretrained NAC encoder. Unlike prior NAC-based SE methods, which process discrete speech tokens using Language Models (LMs), we perform SE within the continuous embedding space of the pretrained NAC, which is highly compressed along the time dimension for efficient representation. Our lightweight SE model, optimized through an embedding-level loss, delivers results comparable to SE baselines trained on larger datasets, with a significantly lower real-time factor of 0.005. Additionally, our method achieves a low GMAC of 3.94, reducing complexity 18-fold compared to Sepformer in a simulated cloud-based audio transmission environment. This work highlights a new, efficient NAC-based SE solution, particularly suitable for cloud applications where NAC is used to compress audio before transmission.

This paper extends Whisper’s automatic speech recognition (ASR) capabilities to perform speech-based emotion recognition (SER) by incorporating word-level emotion classification alongside ASR output. We generate four emotion pseudo-labels (neutral, happy, sad, angry) for each word using a pretrained frame-level SER model, and Whisper is fine-tuned for joint ASR and emotion classification at the word level. Sentence-level emotion labels are masked during training to encourage the transformer to use the ASR output for word-level emotion prediction. During inference, word-level predictions are combined with sentence-level predictions through majority voting to generate the final sentence-level label. When evaluated on the IEMOCAP dataset, our method maintains Whisper’s ASR word error rate while improving the SER weighted accuracy from 74.4% to 76.4% and the unweighted average recall from 77.1% to 79.0%.

Model ensembles using linear interpolation are commonly employed to improve classification performance, with higher weights assigned to better-performing models in the ensemble. However, prior methods use fixed weights across all test samples, which is suboptimal as different models may perform better in different subsets of the samples, especially in out-of-domain (OOD) scenarios. This is a key challenge in Audio Deepfake Detection (ADD) due to variations between training and testing domains. To address this, we propose using EOW-Softmax, a method for modeling open-world uncertainties, to calibrate the magnitudes of OOD classification scores at the sample level. This dynamic adjustment improves ensemble predictions on OOD samples. When tested on the ASVspoof 2021 dataset, our calibrated ensemble reduced the equal error rate (EER) from 2.66% to 2.03%.

The rapid growth of voice assistants powered by large language models (LLM) has highlighted a need for speech instruction data to train these systems. Despite the abundance of speech recognition data, there is a notable scarcity of speech instruction data, which is essential for fine-tuning models to understand and execute spoken commands. Generating high-quality synthetic speech requires a good text-to-speech (TTS) model, which may not be available to low resource languages. Our novel approach addresses this challenge by halting synthesis at the semantic representation level, bypassing the need for TTS. We achieve this by aligning synthetic semantic representations with the pre-trained Whisper encoder, enabling an LLM to be fine-tuned on text instructions while maintaining the ability to understand spoken instructions during inference. This simplified training process is a promising approach to building voice assistant for low-resource languages.

Contextual biasing improves rare word recognition of ASR models by prioritizing the output of rare words during decoding. A common approach is Trie-based biasing, which gives "bonus scores" to partial hypothesis (e.g. "Bon") that may lead to the generation of the rare word (e.g. "Bonham"). If the full word ("Bonham") isn't ultimately recognized, the system revokes those earlier bonuses. This revocation is limited to beam search and is computationally expensive, particularly for models with large decoders. To overcome these limitations, we propose adapting ASR models to look ahead and predict multiple steps at once. This avoids the revocation step entirely by better estimating whether a partial hypothesis will lead to the generation of the full rare word. By fine-tuning Whisper with only 10 hours of synthetic data, our method reduces the word error rate on the NSC Part 2 test set from 30.86% to 12.19%.

Rare word recognition can be improved by adapting ASR models to synthetic data that includes these words. Further improvements can be achieved through contextual biasing, which trains and adds a biasing module into the model architecture to prioritize rare words. While training the module on synthetic rare word data is more effective than using non-rare-word data, it can lead to overfitting due to artifacts in the synthetic audio. To address this, we enhance the TCPGen-based contextual biasing approach and propose a keyword-aware loss function that additionally focuses on biased words when training biasing modules. This loss includes a masked cross-entropy term for biased word prediction and a binary classification term for detecting biased word positions. These two terms complementarily support the decoding of biased words during inference. By adapting Whisper to 10 hours of synthetic data, our method reduced the word error rate on the NSC Part 2 test set from 29.71% to 11.81%.

Audio deepfake detection (ADD) models are commonly evaluated using datasets that combine multiple synthesizers, with performance reported as a single Equal Error Rate (EER). However, this approach disproportionately weights synthesizers with more samples, underrepresenting others and reducing the overall reliability of EER. Additionally, most ADD datasets lack diversity in bona fide speech, often featuring a single environment and speech style (e.g., clean read speech), limiting their ability to simulate real-world conditions. To address these challenges, we propose bona fide cross-testing, a novel evaluation framework that incorporates diverse bona fide datasets and aggregates EERs for more balanced assessments. Our approach improves robustness and interpretability compared to traditional evaluation methods. We benchmark over 150 synthesizers across nine bona fide speech types and release a new dataset to facilitate further research at https://github.com/cyaaronk/audio_deepfake_eval.

Knowledge distillation (KD) is used to enhance automatic speaker verification performance by ensuring consistency between large teacher networks and lightweight student networks at the embedding level or label level. However, the conventional label-level KD overlooks the significant knowledge from non-target speakers, particularly their classification probabilities, which can be crucial for automatic speaker verification. In this paper, we first demonstrate that leveraging a larger number of training non-target speakers improves the performance of automatic speaker verification models. Inspired by this finding about the importance of non-target speakers’ knowledge, we modified the conventional label-level KD by disentangling and emphasizing the classification probabilities of non-target speakers during knowledge distillation. The proposed method is applied to three different student model architectures and achieves an average of 13.67% improvement in EER on the VoxCeleb dataset compared to embedding-level and conventional label-level KD methods.1

Dual-path is a popular architecture for speech separation models (e.g. Sepformer) which splits long sequences into overlapping chunks for its intra- and inter-blocks that separately model intra-chunk local features and inter-chunk global relationships. However, it has been found that inter-blocks, which comprise half a dual-path model’s parameters, contribute minimally to performance. Thus, we propose the Single-Path Global Modulation (SPGM) block to replace inter-blocks. SPGM is named after its structure consisting of a parameter-free global pooling module followed by a modulation module comprising only 2% of the model’s total parameters. The SPGM block allows all transformer layers in the model to be dedicated to local feature modelling, making the overall model single-path. SPGM achieves 22.1 dB SI-SDRi on WSJ0-2Mix and 20.4 dB SI-SDRi on Libri2Mix, exceeding the performance of Sepformer by 0.5 dB and 0.3 dB respectively and matches the performance of recent SOTA models with up to 8 times fewer parameters. Model and weights are available at huggingface.co/yipjiaqi/spgm

Our previously proposed MossFormer has achieved promising performance in monaural speech separation. However, it predominantly adopts a self-attention-based MossFormer module, which tends to emphasize longer-range, coarser-scale dependencies, with a deficiency in effectively modelling finer-scale recurrent patterns. In this paper, we introduce a novel hybrid model that provides the capabilities to model both long-range, coarse-scale dependencies and fine-scale recurrent patterns by integrating a recurrent module into the MossFormer framework. Instead of applying the recurrent neural networks (RNNs) that use traditional recurrent connections, we present a recurrent module based on a feedforward sequential memory network (FSMN), which is considered "RNN-free" recurrent network due to the ability to capture recurrent patterns without using recurrent connections. Our recurrent module mainly comprises an enhanced dilated FSMN block by using gated convolutional units (GCU) and dense connections. In addition, a bottleneck layer and an output layer are also added for controlling information flow. The recurrent module relies on linear projections and convolutions for seamless, parallel processing of the entire sequence. The integrated MossFormer2 hybrid model demonstrates remarkable enhancements over MossFormer and surpasses other state-of-the-art methods in WSJ0-2/3mix, Libri2Mix, and WHAM!/WHAMR! benchmarks.

Neural codecs have become crucial to recent speech and audio generation research. In addition to signal compression capabilities, discrete codecs have also been found to enhance downstream training efficiency and compatibility with autoregressive language models. However, as extensive downstream applications are investigated, challenges have arisen in ensuring fair comparisons across diverse applications. To address these issues, we present a new open-source platform ESPnet-Codec, which is built on ESPnet and focuses on neural codec training and evaluation. ESPnet-Codec offers various recipes in audio, music, and speech for training and evaluation using several widely adopted codec models. Together with ESPnet-Codec, we present VERSA, a standalone evaluation toolkit, which provides a comprehensive evaluation of codec performance over 20 audio evaluation metrics. Notably, we demonstrate that ESPnet-Codec can be integrated into six ESPnet tasks, supporting diverse applications.

Current Multilingual ASR models only support a fraction of the world’s languages. Continual Learning (CL) aims to tackle this problem by adding new languages to pre-trained models while avoiding the loss of performance on existing languages, also known as Catastrophic Forgetting (CF). However, existing CL methods overlook the adaptation of the token embedding lookup table at the decoder, despite its significant contribution to CF. We propose Embedding Layer Surgery where separate copies of the token embeddings are created for each new languages, and one of the copies is selected to replace the old languages embeddings when transcribing the corresponding new language. Unfortunately, this approach means LID errors also cause incorrect ASR embedding selection. Our Task-wise Beam Search allows self-correction for such mistakes. By adapting Whisper to 10 hours of data for each of 10 unseen languages from Common Voice, results show that our method reduces the Average WER (AWER) of pre-trained languages from 14.2% to 11.9% compared with Experience Replay, without compromising the AWER of the unseen languages.

Existing self-supervised pre-trained speech models have offered an effective way to leverage massive unannotated corpora to build good automatic speech recognition (ASR). However, many current models are trained on a clean corpus from a single source, which tends to do poorly when noise is present during testing. Nonetheless, it is crucial to overcome the adverse influence of noise for real-world applications. In this work, we propose a novel training framework, called deHuBERT, for noise reduction encoding inspired by H. Barlow’s redundancy-reduction principle. The new framework improves the HuBERT training algorithm by introducing auxiliary losses that drive the self- and cross-correlation matrix between pairwise noise-distorted embeddings towards identity matrix. This encourages the model to produce noise- agnostic speech representations. With this method, we report improved robustness in noisy environments, including unseen noises, without impairing the performance on the clean set.

Most of the existing neural-based models for keyword spotting (KWS) in smart devices require thousands of training samples to learn a decent audio representation. However, with the rising demand for smart devices to become more person-alized, KWS models need to adapt quickly to smaller user samples. To tackle this challenge, we propose a contrastive speech mixup (CosMix) learning algorithm for low-resource KWS. CosMix introduces an auxiliary contrastive loss to the existing mixup augmentation technique to maximize the relative similarity between the original pre-mixed samples and the augmented samples. The goal is to inject enhancing constraints to guide the model towards simpler but richer content-based speech representations from two augmented views (i.e. noisy mixed and clean pre-mixed utterances). We conduct our experiments on the Google Speech Command dataset, where we trim the size of the training set to as small as 2.5 mins per keyword to simulate a low-resource condition. Our experimental results show a consistent improvement in the performance of multiple models, which exhibits the effectiveness of our method.