Wav2Small: Distilling Wav2Vec2 to 72K parameters for Low-Resource Speech emotion recognition

Speech Emotion Recognition (SER) needs high computational resources to overcome the challenge of substantial annotator disagreement. Today SER is shifting towards dimensional annotations of arousal, dominance, and valence (A/D/V). Universal metrics as the L2 distance prove unsuitable for evaluating A/D/V accuracy due to non converging consensus of annotator opinions. However, Concordance Correlation Coefficient (CCC) arose as an alternative metric for A/D/V where a model's output is evaluated to match a whole dataset's CCC rather than L2 distances of individual audios. Recent studies have shown that Wav2Vec2.0 / WavLM architectures outputing a float value for each A/D/V dimension achieve today's State-of-the-art (SOTA) CCC on A/D/V. The Wav2Vec2.0 / WavLM family has high computational footprint, but training tiny models using human annotations has been unsuccessful. In this paper we use a large Transformer SOTA A/D/V model as Teacher/Annotator to train 5 student models: 4 MobileNets and our proposed Wav2Small, using only the Teacher's A/D/V predictions instead of human annotations. We chose MobileNet-V4 / MobileNet-V3 as students, as MobileNet has been designed for fast execution times. We propose Wav2Small an architecture designed for minimal parameter number and RAM consumption. Wav2Small with an .onnx (quantized) of only $60KB$ is a potential solution for A/D/V on hearing aids, having only 72K parameters vs 3.12M parameters for MobileNet-V4-Small. The Teacher model we construct sets a new SOTA on the MSP Podcast Test-1 dataset with valence CCC=0.676.

Testing Speech Emotion Recognition Machine Learning Models

Machine learning models for speech emotion recognition (SER) can be trained for different tasks and are usually evaluated on the basis of a few available datasets per task. Tasks could include arousal, valence, dominance, emotional categories, or tone of voice. Those models are mainly evaluated in terms of correlation or recall, and always show some errors in their predictions. The errors manifest themselves in model behaviour, which can be very different along different dimensions even if the same recall or correlation is achieved by the model. This paper investigates behavior of speech emotion recognition models with a testing framework which requires models to fulfill conditions in terms of correctness, fairness, and robustness.

Speech-based Age and Gender Prediction with Transformers

We report on the curation of several publicly available datasets for age and gender prediction. Furthermore, we present experiments to predict age and gender with models based on a pre-trained wav2vec 2.0. Depending on the dataset, we achieve an MAE between 7.1 years and 10.8 years for age, and at least 91.1% ACC for gender (female, male, child). Compared to a modelling approach built on handcrafted features, our proposed system shows an improvement of 9% UAR for age and 4% UAR for gender. To make our findings reproducible, we release the best performing model to the community as well as the sample lists of the data splits.

audb -- Sharing and Versioning of Audio and Annotation Data in Python

Driven by the need for larger and more diverse datasets to pre-train and fine-tune increasingly complex machine learning models, the number of datasets is rapidly growing. audb is an open-source Python library that supports versioning and documentation of audio datasets. It aims to provide a standardized and simple user-interface to publish, maintain, and access the annotations and audio files of a dataset. To efficiently store the data on a server, audb automatically resolves dependencies between versions of a dataset and only uploads newly added or altered files when a new version is published. The library supports partial loading of a dataset and local caching for fast access. audb is a lightweight library and can be interfaced from any machine learning library. It supports the management of datasets on a single PC, within a university or company, or within a whole research community. audb is available at https://github.com/audeering/audb.

Probing Speech Emotion Recognition Transformers for Linguistic Knowledge

Large, pre-trained neural networks consisting of self-attention layers (transformers) have recently achieved state-of-the-art results on several speech emotion recognition (SER) datasets. These models are typically pre-trained in self-supervised manner with the goal to improve automatic speech recognition performance -- and thus, to understand linguistic information. In this work, we investigate the extent in which this information is exploited during SER fine-tuning. Using a reproducible methodology based on open-source tools, we synthesise prosodically neutral speech utterances while varying the sentiment of the text. Valence predictions of the transformer model are very reactive to positive and negative sentiment content, as well as negations, but not to intensifiers or reducers, while none of those linguistic features impact arousal or dominance. These findings show that transformers can successfully leverage linguistic information to improve their valence predictions, and that linguistic analysis should be included in their testing.

Dawn of the transformer era in speech emotion recognition: closing the valence gap

Recent advances in transformer-based architectures which are pre-trained in self-supervised manner have shown great promise in several machine learning tasks. In the audio domain, such architectures have also been successfully utilised in the field of speech emotion recognition (SER). However, existing works have not evaluated the influence of model size and pre-training data on downstream performance, and have shown limited attention to generalisation, robustness, fairness, and efficiency. The present contribution conducts a thorough analysis of these aspects on several pre-trained variants of wav2vec 2.0 and HuBERT that we fine-tuned on the dimensions arousal, dominance, and valence of MSP-Podcast, while additionally using IEMOCAP and MOSI to test cross-corpus generalisation. To the best of our knowledge, we obtain the top performance for valence prediction without use of explicit linguistic information, with a concordance correlation coefficient (CCC) of .638 on MSP-Podcast. Furthermore, our investigations reveal that transformer-based architectures are more robust to small perturbations compared to a CNN-based baseline and fair with respect to biological sex groups, but not towards individual speakers. Finally, we are the first to show that their extraordinary success on valence is based on implicit linguistic information learnt during fine-tuning of the transformer layers, which explains why they perform on-par with recent multimodal approaches that explicitly utilise textual information. Our findings collectively paint the following picture: transformer-based architectures constitute the new state-of-the-art in SER, but further advances are needed to mitigate remaining robustness and individual speaker issues. To make our findings reproducible, we release the best performing model to the community.

Referenceless Performance Evaluation of Audio Source Separation using Deep Neural Networks

Current performance evaluation for audio source separation depends on comparing the processed or separated signals with reference signals. Therefore, common performance evaluation toolkits are not applicable to real-world situations where the ground truth audio is unavailable. In this paper, we propose a performance evaluation technique that does not require reference signals in order to assess separation quality. The proposed technique uses a deep neural network (DNN) to map the processed audio into its quality score. Our experiment results show that the DNN is capable of predicting the sources-to-artifacts ratio from the blind source separation evaluation toolkit without the need for reference signals.

Multi-Resolution Fully Convolutional Neural Networks for Monaural Audio Source Separation

In deep neural networks with convolutional layers, each layer typically has fixed-size/single-resolution receptive field (RF). Convolutional layers with a large RF capture global information from the input features, while layers with small RF size capture local details with high resolution from the input features. In this work, we introduce novel deep multi-resolution fully convolutional neural networks (MR-FCNN), where each layer has different RF sizes to extract multi-resolution features that capture the global and local details information from its input features. The proposed MR-FCNN is applied to separate a target audio source from a mixture of many audio sources. Experimental results show that using MR-FCNN improves the performance compared to feedforward deep neural networks (DNNs) and single resolution deep fully convolutional neural networks (FCNNs) on the audio source separation problem.