160 datasets found

Federated Noisy Student Training for Automatic Speech Recognition

The VCTK dataset is a large corpus of speech recordings, each containing a single speaker and a single sentence.

Human speech data comprises a rich set of domain factors such as accent, syntactic and semantic variety, or acoustic environment.

Four different corpora are used for our experiments, which are TIMIT, Aurora-4, AMI, and LibriSpeech. TIMIT contains broadband 16kHz recordings of phonetically-balanced read...

The dataset used in this paper is the BABEL dataset, which contains 10881 motion sequences, with 65926 subsequences and the corresponding textual labels.

The Librispeech dataset is a large-scale speaker-dependent speech corpus containing 1080 hours of speech, 5600 utterances, and 1000 speakers.

The dataset used in this paper is a large-scale production ASR system, which includes multi-domain (MD) data sets in English. The MD data sets include medium-form (MF) and...

The CREMA-D dataset is an audio-visual dataset for emotion recognition task, each video in which consists of both facial and acoustic emotional expressions.

The dataset used in the paper is MSR-VTT, a large video description dataset for bridging video and language. The dataset contains 10k video clips with length varying from 10 to...

Generating holistic 3D human motion from speech.

Speaker verification systems experience significant performance degradation when tasked with short-duration trial recordings. To address this challenge, a multi-scale feature...

Optimization of modern ASR architectures is among the highest priority tasks since it saves many computational resources for model training and inference. The work proposes a...

Voice conversion (VC) is a technique that alters the voice of a source speaker to a target style, such as speaker identity, prosody, and emotion, while keeping the linguistic...

Speaker recognition aims to identify speaker information from input speech. A type of speaker recognition is speaker verification (SV). It determines whether the test speaker's...

The AudioMNIST dataset consists of 60 speakers, 33% female, who were recorded speaking individual digits (0-9) 50 times each.

The Google Speech Commands Dataset Version II contains 105,829 utterances of 35 words from 2,618 speakers with a sampling rate of 16 kHz.

TED-LIUM dataset

The dataset used in the paper is the LibriSpeech dataset, which contains about 1,000 hours of English speech derived from audiobooks.

Deep Speech model

A popular text-based VC approach is to use an automatic speech recognition (ASR) model to extract phonetic posteriorgram (PPG) as content representation.