The VoiceMOS Challenge 2023: Zero-shot Subjective Speech Quality Prediction for Multiple Domains

We present the second edition of the VoiceMOS Challenge, a scientific event that aims to promote the study of automatic prediction of the mean opinion score (MOS) of synthesized and processed speech. This year, we emphasize real-world and challenging zero-shot out-of-domain MOS prediction with three tracks for three different voice evaluation scenarios. Ten teams from industry and academia in seven different countries participated. Surprisingly, we found that the two sub-tracks of French text-to-speech synthesis had large differences in their predictability, and that singing voice-converted samples were not as difficult to predict as we had expected. Use of diverse datasets and listener information during training appeared to be successful approaches.

Improving severity preservation of healthy-to-pathological voice conversion with global style tokens

In healthy-to-pathological voice conversion (H2P-VC), healthy speech is converted into pathological while preserving the identity. The paper improves on previous two-stage approach to H2P-VC where (1) speech is created first with the appropriate severity, (2) then the speaker identity of the voice is converted while preserving the severity of the voice. Specifically, we propose improvements to (2) by using phonetic posteriorgrams (PPG) and global style tokens (GST). Furthermore, we present a new dataset that contains parallel recordings of pathological and healthy speakers with the same identity which allows more precise evaluation. Listening tests by expert listeners show that the framework preserves severity of the source sample, while modelling target speaker's voice. We also show that (a) pathology impacts x-vectors but not all speaker information is lost, (b) choosing source speakers based on severity labels alone is insufficient.

Electrolaryngeal Speech Intelligibility Enhancement Through Robust Linguistic Encoders

We propose a novel framework for electrolaryngeal speech intelligibility enhancement through the use of robust linguistic encoders. Pretraining and fine-tuning approaches have proven to work well in this task, but in most cases, various mismatches, such as the speech type mismatch (electrolaryngeal vs. typical) or a speaker mismatch between the datasets used in each stage, can deteriorate the conversion performance of this framework. To resolve this issue, we propose a linguistic encoder robust enough to project both EL and typical speech in the same latent space, while still being able to extract accurate linguistic information, creating a unified representation to reduce the speech type mismatch. Furthermore, we introduce HuBERT output features to the proposed framework for reducing the speaker mismatch, making it possible to effectively use a large-scale parallel dataset during pretraining. We show that compared to the conventional framework using mel-spectrogram input and output features, using the proposed framework enables the model to synthesize more intelligible and naturally sounding speech, as shown by a significant 16% improvement in character error rate and 0.83 improvement in naturalness score.

AAS-VC: On the Generalization Ability of Automatic Alignment Search based Non-autoregressive Sequence-to-sequence Voice Conversion

Non-autoregressive (non-AR) sequence-to-seqeunce (seq2seq) models for voice conversion (VC) is attractive in its ability to effectively model the temporal structure while enjoying boosted intelligibility and fast inference thanks to non-AR modeling. However, the dependency of current non-AR seq2seq VC models on ground truth durations extracted from an external AR model greatly limits its generalization ability to smaller training datasets. In this paper, we first demonstrate the above-mentioned problem by varying the training data size. Then, we present AAS-VC, a non-AR seq2seq VC model based on automatic alignment search (AAS), which removes the dependency on external durations and serves as a proper inductive bias to provide the required generalization ability for small datasets. Experimental results show that AAS-VC can generalize better to a training dataset of only 5 minutes. We also conducted ablation studies to justify several model design choices. The audio samples and implementation are available online.

Audio Difference Learning for Audio Captioning

This study introduces a novel training paradigm, audio difference learning, for improving audio captioning. The fundamental concept of the proposed learning method is to create a feature representation space that preserves the relationship between audio, enabling the generation of captions that detail intricate audio information. This method employs a reference audio along with the input audio, both of which are transformed into feature representations via a shared encoder. Captions are then generated from these differential features to describe their differences. Furthermore, a unique technique is proposed that involves mixing the input audio with additional audio, and using the additional audio as a reference. This results in the difference between the mixed audio and the reference audio reverting back to the original input audio. This allows the original input's caption to be used as the caption for their difference, eliminating the need for additional annotations for the differences. In the experiments using the Clotho and ESC50 datasets, the proposed method demonstrated an improvement in the SPIDEr score by 7% compared to conventional methods.

Evaluating Methods for Ground-Truth-Free Foreign Accent Conversion

Foreign accent conversion (FAC) is a special application of voice conversion (VC) which aims to convert the accented speech of a non-native speaker to a native-sounding speech with the same speaker identity. FAC is difficult since the native speech from the desired non-native speaker to be used as the training target is impossible to collect. In this work, we evaluate three recently proposed methods for ground-truth-free FAC, where all of them aim to harness the power of sequence-to-sequence (seq2seq) and non-parallel VC models to properly convert the accent and control the speaker identity. Our experimental evaluation results show that no single method was significantly better than the others in all evaluation axes, which is in contrast to conclusions drawn in previous studies. We also explain the effectiveness of these methods with the training input and output of the seq2seq model and examine the design choice of the non-parallel VC model, and show that intelligibility measures such as word error rates do not correlate well with subjective accentedness. Finally, our implementation is open-sourced to promote reproducible research and help future researchers improve upon the compared systems.

Preference-based training framework for automatic speech quality assessment using deep neural network

One objective of Speech Quality Assessment (SQA) is to estimate the ranks of synthetic speech systems. However, recent SQA models are typically trained using low-precision direct scores such as mean opinion scores (MOS) as the training objective, which is not straightforward to estimate ranking. Although it is effective for predicting quality scores of individual sentences, this approach does not account for speech and system preferences when ranking multiple systems. We propose a training framework of SQA models that can be trained with only preference scores derived from pairs of MOS to improve ranking prediction. Our experiment reveals conditions where our framework works the best in terms of pair generation, aggregation functions to derive system score from utterance preferences, and threshold functions to determine preference from a pair of MOS. Our results demonstrate that our proposed method significantly outperforms the baseline model in Spearman's Rank Correlation Coefficient.

The Singing Voice Conversion Challenge 2023

We present the latest iteration of the voice conversion challenge (VCC) series, a bi-annual scientific event aiming to compare and understand different voice conversion (VC) systems based on a common dataset. This year we shifted our focus to singing voice conversion (SVC), thus named the challenge the Singing Voice Conversion Challenge (SVCC). A new database was constructed for two tasks, namely in-domain and cross-domain SVC. The challenge was run for two months, and in total we received 26 submissions, including 2 baselines. Through a large-scale crowd-sourced listening test, we observed that for both tasks, although human-level naturalness was achieved by the top system, no team was able to obtain a similarity score as high as the target speakers. Also, as expected, cross-domain SVC is harder than in-domain SVC, especially in the similarity aspect. We also investigated whether existing objective measurements were able to predict perceptual performance, and found that only few of them could reach a significant correlation.

An Analysis of Personalized Speech Recognition System Development for the Deaf and Hard-of-Hearing

Deaf or hard-of-hearing (DHH) speakers typically have atypical speech caused by deafness. With the growing support of speech-based devices and software applications, more work needs to be done to make these devices inclusive to everyone. To do so, we analyze the use of openly-available automatic speech recognition (ASR) tools with a DHH Japanese speaker dataset. As these out-of-the-box ASR models typically do not perform well on DHH speech, we provide a thorough analysis of creating personalized ASR systems. We collected a large DHH speaker dataset of four speakers totaling around 28.05 hours and thoroughly analyzed the performance of different training frameworks by varying the training data sizes. Our findings show that 1000 utterances (or 1-2 hours) from a target speaker can already significantly improve the model performance with minimal amount of work needed, thus we recommend researchers to collect at least 1000 utterances to make an efficient personalized ASR system. In cases where 1000 utterances is difficult to collect, we also discover significant improvements in using previously proposed data augmentation techniques such as intermediate fine-tuning when only 200 utterances are available.

Text-to-speech synthesis based on latent variable conversion using diffusion probabilistic model and variational autoencoder

Text-to-speech synthesis (TTS) is a task to convert texts into speech. Two of the factors that have been driving TTS are the advancements of probabilistic models and latent representation learning. We propose a TTS method based on latent variable conversion using a diffusion probabilistic model and the variational autoencoder (VAE). In our TTS method, we use a waveform model based on VAE, a diffusion model that predicts the distribution of latent variables in the waveform model from texts, and an alignment model that learns alignments between the text and speech latent sequences. Our method integrates diffusion with VAE by modeling both mean and variance parameters with diffusion, where the target distribution is determined by approximation from VAE. This latent variable conversion framework potentially enables us to flexibly incorporate various latent feature extractors. Our experiments show that our method is robust to linguistic labels with poor orthography and alignment errors.