Interactions with virtual assistants typically start with a predefined trigger phrase followed by the user command. To make interactions with the assistant more intuitive, we explore whether it is feasible to drop the requirement that users must begin each command with a trigger phrase. We explore this task in three ways: First, we train classifiers using only acoustic information obtained from the audio waveform. Second, we take the decoder outputs of an automatic speech recognition (ASR) system, such as 1-best hypotheses, as input features to a large language model (LLM). Finally, we explore a multimodal system that combines acoustic and lexical features, as well as ASR decoder signals in an LLM. Using multimodal information yields relative equal-error-rate improvements over text-only and audio-only models of up to 39% and 61%. Increasing the size of the LLM and training with low-rank adaption leads to further relative EER reductions of up to 18% on our dataset.
It is well-known that the reparameterisation gradient estimator, which exhibits low variance in practice, is biased for non-differentiable models. This may compromise correctness of gradient-based optimisation methods such as stochastic gradient descent (SGD). We introduce a simple syntactic framework to define non-differentiable functions piecewisely and present a systematic approach to obtain smoothings for which the reparameterisation gradient estimator is unbiased. Our main contribution is a novel variant of SGD, Diagonalisation Stochastic Gradient Descent, which progressively enhances the accuracy of the smoothed approximation during optimisation, and we prove convergence to stationary points of the unsmoothed (original) objective. Our empirical evaluation reveals benefits over the state of the art: our approach is simple, fast, stable and attains orders of magnitude reduction in work-normalised variance.
Interactions with virtual assistants typically start with a trigger phrase followed by a command. In this work, we explore the possibility of making these interactions more natural by eliminating the need for a trigger phrase. Our goal is to determine whether a user addressed the virtual assistant based on signals obtained from the streaming audio recorded by the device microphone. We address this task by combining 1-best hypotheses and decoder signals from an automatic speech recognition system with acoustic representations from an audio encoder as input features to a large language model (LLM). In particular, we are interested in data and resource efficient systems that require only a small amount of training data and can operate in scenarios with only a single frozen LLM available on a device. For this reason, our model is trained on 80k or less examples of multimodal data using a combination of low-rank adaptation and prefix tuning. We compare the proposed system to unimodal baselines and show that the multimodal approach achieves lower equal-error-rates (EERs), while using only a fraction of the training data. We also show that low-dimensional specialized audio representations lead to lower EERs than high-dimensional general audio representations.
Cycle-consistent generative adversarial networks have been widely used in non-parallel voice conversion (VC). Their ability to learn mappings between source and target features without relying on parallel training data eliminates the need for temporal alignments. However, most methods decouple the conversion of acoustic features from synthesizing the audio signal by using separate models for conversion and waveform synthesis. This work unifies conversion and synthesis into a single model, thereby eliminating the need for a separate vocoder. By leveraging cycle-consistent training and a self-supervised auxiliary training task, our model is able to efficiently generate converted high-quality raw audio waveforms. Subjective listening tests show that our method outperforms the baseline in whispered speech conversion (up to 6.7% relative improvement), and mean opinion score predictions yield competitive results in conventional VC (between 0.5% and 2.4% relative improvement).
Most stuttering detection and classification research has viewed stuttering as a multi-class classification problem or a binary detection task for each dysfluency type; however, this does not match the nature of stuttering, in which one dysfluency seldom comes alone but rather co-occurs with others. This paper explores multi-language and cross-corpus end-to-end stuttering detection as a multi-label problem using a modified wav2vec 2.0 system with an attention-based classification head and multi-task learning. We evaluate the method using combinations of three datasets containing English and German stuttered speech, one containing speech modified by fluency shaping. The experimental results and an error analysis show that multi-label stuttering detection systems trained on cross-corpus and multi-language data achieve competitive results but performance on samples with multiple labels stays below over-all detection results.
We study the foundations of variational inference, which frames posterior inference as an optimisation problem, for probabilistic programming. The dominant approach for optimisation in practice is stochastic gradient descent. In particular, a variant using the so-called reparameterisation gradient estimator exhibits fast convergence in a traditional statistics setting. Unfortunately, discontinuities, which are readily expressible in programming languages, can compromise the correctness of this approach. We consider a simple (higher-order, probabilistic) programming language with conditionals, and we endow our language with both a measurable and a smoothed (approximate) value semantics. We present type systems which establish technical pre-conditions. Thus we can prove stochastic gradient descent with the reparameterisation gradient estimator to be correct when applied to the smoothed problem. Besides, we can solve the original problem up to any error tolerance by choosing an accuracy coefficient suitably. Empirically we demonstrate that our approach has a similar convergence as a key competitor, but is simpler, faster, and attains orders of magnitude reduction in work-normalised variance.
This work adapts two recent architectures of generative models and evaluates their effectiveness for the conversion of whispered speech to normal speech. We incorporate the normal target speech into the training criterion of vector-quantized variational autoencoders (VQ-VAEs) and MelGANs, thereby conditioning the systems to recover voiced speech from whispered inputs. Objective and subjective quality measures indicate that both VQ-VAEs and MelGANs can be modified to perform the conversion task. We find that the proposed approaches significantly improve the Mel cepstral distortion (MCD) metric by at least 25% relative to a DiscoGAN baseline. Subjective listening tests suggest that the MelGAN-based system significantly improves naturalness, intelligibility, and voicing compared to the whispered input speech. A novel evaluation measure based on differences between latent speech representations also indicates that our MelGAN-based approach yields improvements relative to the baseline.
We analyze the impact of speaker adaptation in end-to-end architectures based on transformers and wav2vec 2.0 under different noise conditions. We demonstrate that the proven method of concatenating speaker vectors to the acoustic features and supplying them as an auxiliary model input remains a viable option to increase the robustness of end-to-end architectures. By including speaker embeddings obtained from x-vector and ECAPA-TDNN models, we achieve relative word error rate improvements of up to 9.6% on LibriSpeech and up to 14.5% on Switchboard. The effect on transformer-based architectures is approximately inversely proportional to the signal-to-noise ratio (SNR) and is strongest in heavily noised environments ($SNR=0$). The most substantial benefit of speaker adaption in systems based on wav2vec 2.0 can be achieved under moderate noise conditions ($SNR\geq18$). We also find that x-vectors tend to yield larger improvements than ECAPA-TDNN embeddings.
Specially adapted speech recognition models are necessary to handle stuttered speech. For these to be used in a targeted manner, stuttered speech must be reliably detected. Recent works have treated stuttering as a multi-class classification problem or viewed detecting each dysfluency type as an isolated task; that does not capture the nature of stuttering, where one dysfluency seldom comes alone, i.e., co-occurs with others. This work explores an approach based on a modified wav2vec 2.0 system for end-to-end stuttering detection and classification as a multi-label problem. The method is evaluated on combinations of three datasets containing English and German stuttered speech, yielding state-of-the-art results for stuttering detection on the SEP-28k-Extended dataset. Experimental results provide evidence for the transferability of features and the generalizability of the method across datasets and languages.
Current findings show that pre-trained wav2vec 2.0 models can be successfully used as feature extractors to discriminate on speaker-based tasks. We demonstrate that latent representations extracted at different layers of a pre-trained wav2vec 2.0 system can be effectively used for binary classification of various types of pathologic speech. We examine the pathologies laryngectomy, oral squamous cell carcinoma, parkinson's disease and cleft lip and palate for this purpose. The results show that a distinction between pathological and healthy voices, especially with latent representations from the lower layers, performs well with the lowest accuracy from 77.2% for parkinson's disease to 100% for laryngectomy classification. However, cross-pathology and cross-healthy tests show that the trained classifiers seem to be biased. The recognition rates vary considerably if there is a mismatch between training and out-of-domain test data, e.g., in age, spoken content or acoustic conditions.