LSTM-based speaker verification usually uses a fixed-length local segment randomly truncated from an utterance to learn the utterance-level speaker embedding, while using the average embedding of all segments of a test utterance to verify the speaker, which results in a critical mismatch between testing and training. T…
On-device research index
arXiv research
A locally-built, LLM-digested index of recent arXiv papers in quant finance, geometry/topology, and statistical ML — keyword search served straight from SQLite on this machine.
168,932 papers · 148 categories
Trend · papers per month
48 results for “utterance embeddings”
Improves speaker verification for variable-duration utterances using a feature pyramid module.
problem Improving robustness for variable-duration utterances in speaker verification.
method Integrates a feature pyramid module into multi-scale aggregation to enhance speaker-discriminative information from multiple layers.
result Improves performance for both short and long utterances compared to state-of-the-art approaches.
Simple baselines improve multimodal utterance learning.
problem Learning rich multimodal utterance representations.
method Conditional factorization of utterances into unimodal factors; extending to bimodal and trimodal factors.
result Optimal embeddings can be derived in closed form.
Improved far-field speaker verification for short utterances in noisy conditions.
problem Challenges in speaker verification on short utterances in uncontrolled noisy environments.
method Used deep neural network architectures (TDNN and ResNet) and experimented with various embedding extractors and training procedures.
result ResNet architectures outperform x-vector approach in speaker verification quality for both long and short utterances.
Self multi-head attention improves speaker recognition for long utterances.
problem Speaker recognition for long speech segments using Deep Learning.
method Convolutional Neural Network (CNN) for short-term features, self multi-head attention for long-term embeddings.
result Self multi-head attention outperforms other pooling methods by 18% relative EER on VoxCeleb1 dataset.
Meta-learning framework improves short utterance speaker recognition.
problem Poor performance of existing models with short utterances.
method Prototypical Networks with support and query sets, enforcing classification against entire training set.
result Significant performance gains on VoxCeleb datasets.
User interaction with voice-powered agents generates large amounts of unlabeled utterances. In this paper, we explore techniques to efficiently transfer the knowledge from these unlabeled utterances to improve model performance on Spoken Language Understanding (SLU) tasks. We use Embeddings from Language Model (ELMo) t…
Modeling long-range context for multi-function utterances in dialogues.
problem Complex dependencies across dialogue turns in long utterances.
method Adapted Convolutional Recurrent Neural Network (CRNN) to model interactions between utterances.
result Significantly outperforms existing work on CDA recognition on a tech forum dataset.
Proposes DNN-based speaker embedding correlated with subjective inter-speaker similarity for speech synthesis.
problem Inadequate speaker representation for open speakers not in training data.
method Two training algorithms using inter-speaker similarity matrices: similarity vector embedding and similarity matrix embedding.
result Proposed algorithms learn speaker embedding highly correlated with subjective inter-speaker similarity.
In this paper, we propose a new differentiable neural network alignment mechanism for text-dependent speaker verification which uses alignment models to produce a supervector representation of an utterance. Unlike previous works with similar approaches, we do not extract the embedding of an utterance from the mean redu…
Text-independent speaker recognition using short utterances is a highly challenging task due to the large variation and content mismatch between short utterances. I-vector based systems have become the standard in speaker verification applications, but they are less effective with short utterances. In this paper, we fi…
GE2E-AC improves accent classification by focusing on accent embeddings.
problem Training models to predict accent type can lead to learning irrelevant features.
method GE2E-AC trains models to extract accent embeddings, making them closer for the same accent class.
result GE2E-AC outperforms baseline models trained with conventional loss.
Graphical lasso models ASR utterance dependencies for consistent WER estimation.
problem Modeling dependent structure among ASR utterances for accurate significance analysis.
method Graphical lasso for dependency modeling, followed by blockwise bootstrap resampling.
result Statistically consistent variance estimator of WER under mild conditions.
Speaker embedding models that utilize neural networks to map utterances to a space where distances reflect similarity between speakers have driven recent progress in the speaker recognition task. However, there is still a significant performance gap between recognizing speakers in the training set and unseen speakers. …
The recent surge of text-based online counseling applications enables us to collect and analyze interactions between counselors and clients. A dataset of those interactions can be used to learn to automatically classify the client utterances into categories that help counselors in diagnosing client status and predictin…
Speaker verification (SV) systems using deep neural network embeddings, so-called the x-vector systems, are becoming popular due to its good performance superior to the i-vector systems. The fusion of these systems provides improved performance benefiting both from the discriminatively trained x-vectors and generative …
DGP learns speech recognition by modeling complex relationships between utterances.
problem Modeling complex relationships in speech recognition without relational data.
method Bayesian nonparametric deep learning method (DGP) that generates infinite probabilistic graphs.
result DGP successfully infers relationships among utterances without relational data during training.
A new type of End-to-End system for text-dependent speaker verification is presented in this paper. Previously, using the phonetically discriminative/speaker discriminative DNNs as feature extractors for speaker verification has shown promising results. The extracted frame-level (DNN bottleneck, posterior or d-vector) …
Improved voice conversion with semi-supervised learning.
problem Voice conversion with limited parallel data.
method Amortized variational inference with parallel and non-parallel utterances.
result Semi-supervised training improves voice conversion performance.
An End-to-End Text-independent Speaker Verification Framework with a Keyword Adversarial Networkeess.AS
End-to-end speaker verification framework reduces text dependency.
problem Improving text-independent speaker verification.
method Jointly trains SE and ASR networks with triplet loss and adversarial gradient.
result Lower equal error rate and better text-independency compared to other approaches.
Recurrent sequence generators conditioned on input data through an attention mechanism have recently shown very good performance on a range of tasks in- cluding machine translation, handwriting synthesis and image caption gen- eration. We extend the attention-mechanism with features needed for speech recognition. We sh…
Speech enhancement model is used to map a noisy speech to a clean speech. In the training stage, an objective function is often adopted to optimize the model parameters. However, in most studies, there is an inconsistency between the model optimization criterion and the evaluation criterion on the enhanced speech. For …
Activates speech DNNs to generate understandable examples.
problem Difficulty in understanding DNN classifications for speech.
method Activation maximization to generate speech samples.
result Activation maximization can generate understandable speech samples.
This paper presents sampling-based speech parameter generation using moment-matching networks for Deep Neural Network (DNN)-based speech synthesis. Although people never produce exactly the same speech even if we try to express the same linguistic and para-linguistic information, typical statistical speech synthesis pr…
Extracting Structured Data from Physician-Patient Conversations By Predicting Noteworthy Utterancescs.LG
The paper extracts structured data from physician-patient conversations, reducing clerical burden.
problem Mining insights from physician-patient conversations for electronic health record documentation.
method Created a dataset of transcripts and summaries, extracted noteworthy utterances, and improved model performance.
result Extracting noteworthy utterances significantly boosts model performance for recognizing diagnoses and RoS abnormalities.
Paper proposes ICCN to learn correlations between text, audio, and video for multimodal sentiment analysis.
problem Improving multimodal sentiment analysis by learning hidden correlations between text and audio/video features.
method Interaction Canonical Correlation Network (ICCN) using deep canonical correlation analysis (DCCA).
result Empirical results confirm the effectiveness of ICCN in capturing useful information from all three views.
In this paper, we propose a fully supervised speaker diarization approach, named unbounded interleaved-state recurrent neural networks (UIS-RNN). Given extracted speaker-discriminative embeddings (a.k.a. d-vectors) from input utterances, each individual speaker is modeled by a parameter-sharing RNN, while the RNN state…
Developers of text-to-speech synthesizers (TTS) often make use of human raters to assess the quality of synthesized speech. We demonstrate that we can model human raters' mean opinion scores (MOS) of synthesized speech using a deep recurrent neural network whose inputs consist solely of a raw waveform. Our best models …
Deep learning is still not a very common tool in speaker verification field. We study deep convolutional neural network performance in the text-prompted speaker verification task. The prompted passphrase is segmented into word states - i.e. digits -to test each digit utterance separately. We train a single high-level f…
Two large medical dialogue datasets for improving healthcare.
problem Improving healthcare through better dialogue systems.
method Building two large-scale medical dialogue datasets: MedDialog-EN and MedDialog-CN.
result The datasets are the largest medical dialogue datasets to date.
We investigate the performance of features that can capture nonlinear recurrence dynamics embedded in the speech signal for the task of Speech Emotion Recognition (SER). Reconstruction of the phase space of each speech frame and the computation of its respective Recurrence Plot (RP) reveals complex structures which can…
Learning goal-oriented dialogues by means of deep reinforcement learning has recently become a popular research topic. However, commonly used policy-based dialogue agents often end up focusing on simple utterances and suboptimal policies. To mitigate this problem, we propose a class of novel temperature-based extension…
Anonymizes speech data to protect privacy.
problem Protecting personal speech data from misuse.
method Extracts features, uses x-vectors, and neural models to synthesize anonymized speech.
result Effective in concealing speaker identities without compromising speech quality.
We propose an adversarial learning approach for generating multi-turn dialogue responses. Our proposed framework, hredGAN, is based on conditional generative adversarial networks (GANs). The GAN's generator is a modified hierarchical recurrent encoder-decoder network (HRED) and the discriminator is a word-level bidirec…
Inspired by recent work on neural network image generation which rely on backpropagation towards the network inputs, we present a proof-of-concept system for speech texture synthesis and voice conversion based on two mechanisms: approximate inversion of the representation learned by a speech recognition neural network,…
Self-supervised method detects replay spoofing using acoustic configurations.
problem Challenges in collecting large-scale datasets for replay spoofing detection.
method Self-supervised pretraining of acoustic configurations using existing datasets.
result The method outperforms baseline by 30% on ASVspoof 2019 physical access dataset.
Blockwise bootstrap improves ASR performance testing for correlated data.
problem Testing reliability of WER improvements between ASR systems.
method Divide evaluation utterances into nonoverlapping blocks and resample these blocks.
result The variance estimator of absolute WER difference is consistent under mild conditions.
A standard recipe for spoken language recognition is to apply a Gaussian back-end to i-vectors. This ignores the uncertainty in the i-vector extraction, which could be important especially for short utterances. A recent paper by Cumani, Plchot and Fer proposes a solution to propagate that uncertainty into the backend. …
Paper introduces methods to automatically generate SOAP notes from patient-physician conversations.
problem Burden of creating digital SOAP notes by physicians.
method Cluster2Sent algorithm for summarizing patient-physician conversations.
result Cluster2Sent algorithm outperforms existing methods by 8 ROUGE-1 points.
Attention-based encoder decoder network uses a left-to-right beam search algorithm in the inference step. The current beam search expands hypotheses and traverses the expanded hypotheses at the next time step. This traversal is implemented using a for-loop program in general, and it leads to speed down of the recogniti…
Study combines chit-chat and goal-oriented dialogue in fantasy games.
problem Combining naturalistic chit-chat with goal-oriented tasks in fantasy games.
method Trained a goal-oriented model with reinforcement learning against an imitation-learned chit-chat model using two approaches.
result Both models outperform a baseline and can converse naturally to achieve goals.
This paper improves speech synthesis using a DGP with SRU for naturalness.
problem Improving naturalness in synthetic speech.
method Deep Gaussian process with a recurrent architecture using SRU.
result SRU-DGP outperforms other models in naturalness of synthetic speech.
SuTaT creates dialogue summaries for tete-a-tetes without labeled data.
problem Lack of high-quality paired dialogue-summary data.
method Unsupervised model for tete-a-tetes, modeling customer and agent roles separately.
result SuTaT outperforms on automatic and human evaluations.
In this paper, we extend the persona-based sequence-to-sequence (Seq2Seq) neural network conversation model to a multi-turn dialogue scenario by modifying the state-of-the-art hredGAN architecture to simultaneously capture utterance attributes such as speaker identity, dialogue topic, speaker sentiments and so on. The …
HRL improves open-domain dialog models by optimizing long-term conversational goals.
problem Challenges in open-domain dialog generation, including repetitive outputs, difficulty tracking conversational goals, and inappropriate text.
method Proposes VHRL, a hierarchical reinforcement learning approach using policy gradients to tune utterance-level embeddings of a variational sequence model.
result Significant improvements in human evaluation and automatic metrics over state-of-the-art dialog models.
Improves E2E ASR performance on numeric sequences with additional training data and denormalization.
problem Challenges in recognizing numeric sequences out-of-vocabulary in ASR systems.
method Uses text-to-speech for additional numeric training data and a small-footprint neural network for denormalization.
result Reduction of WER by up to a factor of 8 in the longest numeric sequences.
Recent approaches based on artificial neural networks (ANNs) have shown promising results for short-text classification. However, many short texts occur in sequences (e.g., sentences in a document or utterances in a dialog), and most existing ANN-based systems do not leverage the preceding short texts when classifying …
Speech Synthesis using EEGeess.AS
Speech synthesis from EEG features using RNN.
problem Speech synthesis from EEG data.
method Recurrent Neural Network (RNN) regression model to predict acoustic features from EEG features.
result Feasibility of synthesizing speech directly from EEG features demonstrated.