We present the corpus called HealthCall which was recorded in real-life conditions in the call center of Malakoff Humanis a health insurance company. The records include two separate audio channels, the first one for the customer and the second one for the agent. This corpus includes a transcription of the spoken conversations and was divided into three sets: training (Train), development (Dev) and test (Test) sets. Two customer relationship management tasks were assessed on the HealthCall corpus: the classification of user requests and the detection of complaints. For this purpose, we have investigated 18 feature sets: 12 linguistic and 6 audio sets. We have used BERT models for the linguistic features and Wav2Vec models for the audio features. The results show that the linguistic features give always the best results (92.7% for the Request task and 69.0% for the Complaint task) but the concatenation of acoustic and linguistic features allows a slight improvement for the Complaint task (69.4% versus 69%).

The ACM Multimedia 2023 Computational Paralinguistics Challenge addresses two different problems for the first time in a research competition under well-defined conditions: In the Emotion Share Sub-Challenge, a regression on speech has to be made; and in the Requests Sub-Challenges, requests and complaints need to be detected. We describe the Sub-Challenges, baseline feature extraction, and classifiers based on the 'usual' ComPaRE features, the auDeep toolkit, and deep feature extraction from pre-trained CNNs using the DeepSpectRum toolkit; in addition, wav2vec2 models are used.

The ACM Multimedia 2022 Stuttering Challenge is to determine the stuttering-related class of a speech segment. There are seven stuttering-related classes and an eighth garbage class. For this purpose, we have investigated the Wav2Vec 2.0 deep neural network to extract audio features from Transformer embeddings. Experiments were conducted on a part of the Kassel State of Fluency Corpus (KSoF). First, we introduced 21 functionals allowing to define two large composite audio feature: the first one set called W2V2 Basic audio-feature set (193,536 features) from the Base version of Transformer embeddings and the second one called W2V2 Large audio-feature set (516, 096 features) from the Large version of Transformer embeddings. Some functionals aim at estimating the spatial variability (e.g., mean, standard deviation, quartiles) and others aim at the temporal variability (e.g., linear regression slope). Then, each composite audio feature set have been splitted into a set of audio feature sets by grouping audio features by functional and by layer. Then, the most discriminant audio feature sets have been selected from these audio feature sets. Finally, two audio features sets specializing in stuttering speech have been developed and assessed: W2V2 Advanced audio-feature set (9,984 features) and W2V2 Large Advanced audio-feature set (15,360 features). Experiments have shown an improvement of 9.3% to the first one and 11.9% for the second one on the Test set compared to the official baseline of the Challenge (40.4%).

International audience

International audience

International audience

International audience

The Cognitive Load (CL) refers to the load imposed on an individual’s cognitive system when performing a given task, and is usually associated with the limitations of the human working memory. Stress, fatigue, lower ability to make decisions and perceptual narrowing are induced by cognitive overload which occurs when too much information has to be processed. As many physiological measures and for a non-intrusive measurement, speech features have been investigated in order to find reliable indicators of CL levels. In this paper, we have investigated high-level speech events automatically detected using the CMU-Sphinx toolkit for speech recognition. Temporal events (speech onset latency, event starting time-codes, pause and phone segments) were extracted from the speech transcriptions (phoneme, word, silent pause, filled pause, breathing). Seven audio feature sets related to the speech events were designed and assessed. Three-class SVM classifiers (Low, Medium and High level) were developed and assessed on the CSLE (Cognitive-Load with Speech and EGG) databases provided for the Interspeech'2014 Cognitive Load Sub-Challenge. These experiments have shown an improvement of 1.5 % on the Test set compared to the official baseline Unweighted Average Recall (UAR). Index Terms: cognitive load measurement, silent and filler pause, speech onset latency, paralinguistic challenge 1.

TeraSpeech is a bilingual database (i.e., English and French) developed in partnership with a French museum, le Musee des Sciences et de l’Industrie in Paris. A demonstration of vocal signature is the support of this data collection. Aiming at the validation of a quality plan, a scenario of the demonstration has been designed, and various protocols have been developed. The quality plan is presented as well as the solutions we found for its validation (i.e., scenario and protocols). The statistics of TeraSpeech are given. Three trends are examined for the perspectives : the validation, the exploitation and the research. Over a single year of the vocal signature exhibition, TeraSpeech’2000 is a collection of more than 30,000 sentences recorded from more than 10,000 visitors. The exposition on acoustics of the museum is planned for ten years. TeraSpeech is expected to be a collection of more than 100,000 speakers recorded over the same sound acquisition channel.

International audience

International audience

International audience

International audience

In order to improve the performances of speaker recognition on telephone speech, we investigate the ability to cooperate of two different natures modelizations: the GMM and the ARVM.For the cooperation and competition of the GMM and ARVM modelizations, we used normalized measures.We develop two approaches for these cooperation and competition : a global approach and an analytical approach.We investigate experiments on whole sentences or selected phonetic segments.Theses approaches allow us to obtain performances improvements for both cooperation and competition, and good results on 168 speakers of the NTIMIT database (GMM : 61.7 %, ARVM : 78.1 %, cooperation : 79.9 % and competition : 82.6 %).

In this paper, we study various technics to improve the performance, to reduce the computation cost and the required memory of a recognition system based on HMM. For the efficiency of the system, we first study the optimization of the number of HMM parameters according to training data. We experiment a temporal control of the phonetic transitions on lexical decoding task with a significant 5% improvement. Finally, a preliminary method selecting dynamically a sublexicon is studied in order to reduce the lexical decoding cost. 1. INTRODUCTION Hidden Markov Models (HMM) produce the most outstanding results in the field of speech recognition. However, this technic has several limitations such as large amount of computation cost and memory request, and the difficulty to represent the duration in hidden Markov models. For the efficiency of the system, we first sudy the optimization of the number of HMM parameters according to training data. It consists in an iterative procedure to choice ...

In this paper, we compare three speaker recognition systems results (i.e. GMM, AHSM, ARVM) on the TIMIT and NTIMIT databases. In order to improve the results on the NTIMIT database, we present two more sophisticated systems: the first one is based on ARMAVector model, the second one is based on the utilisation of several AR-Vector models per speaker. We investigate the ability to recognize speaker using phonetic segments labels. We test the cooperation of several AR-Vector models, each one being learned on a distinct phonetic cluster. For this cooperation, we develop a segmental and an analytic approaches. 1. INTRODUCTION Speaker recognition refers to two problems which are the verification of the claimed identity and the speaker identification. In this paper, we are especially interested in speaker identification in phone quality speech. This task is of a high interest in vocal applications through the telephone network. We use for our experiments the NTIMIT database [1]. The NTIMIT...

We investigate speaker-related information in speech. Our investigations are based on the AR-vector model used in the Orphee system: the vocal activated door we designed. The free-text speaker identification accuracy is an indicator of the ability of any parameter for discriminating speakers. Therefore, through speaker identification performance, we test the speech duration influence, the effect of the parameters of AR-vector model and the importance of phonetic segments. We developed a method to predict the most speaker-related phonemes. The results are the following: the vowels, the diphtongues and the nasals are shown to be the most specific to speakers discrimination. For one second duration test signal regrouping vowels and diphtongues, a speaker accuracy rate of 95.6% was found for the 630 TIMIT speakers.>

Two models, the temporal decomposition and the multivariate linear prediction, of the spectral evolution of speech signals capable of processing some aspects of the speech variability are presented. A series of acoustic-phonetic decoding experiments, characterized by the use of spectral targets of the temporal decomposition techniques and a speaker-dependent mode, gives good results compared to a reference system (i.e., 70% vs. 60% for the first choice). Using the original method developed by Laforia, a series of text-independent speaker recognition experiments, characterized by a long-term multivariate auto-regressive modelization, gives first-rate results (i.e., 98.4% recognition rate for 420 speakers) without using more than one sentence. Taking into account the interpretation of the models, these results show how interesting the cinematic models are for obtaining a reduced variability of the speech signal representation.>

In this paper, a new text-independent speaker recognition method is proposed. This method uses a modeling of the spectral evolution of the speech signals, which is capable of processing some aspects of the inter-speaker variability : the AR-Vector models. Some inter-speaker measures are presented and their advantages/inconvenients are discussed. A training technique to learn discriminant AR-Vector models is proposed. The evaluation of this method is carried out on the TIMIT database recorded by cooperative speakers without any impostor. A series of text-independent speaker identification experiments are described. There is no specific corpus for the training sentences and the training corpus is different from the test corpus. Two speech qualities are tested (i.e., good quality and phone quality). The experiments with good speech quality give first-rate results (i.e, identification rate of 100% for 420 speakers) without using more than two sentences for each test. I. INTRODUCTION Speak...

A report is presented of intraspeaker and interspeaker variability as a major source of error in automatic speech recognition. The authors report on two series of experiments using multilayer feed-forward automata (MLFFA) to control some aspects of this variability. The first series concerns the classification of spectral targets obtained from a robust implementation of temporal decomposition. An MLFFA accepts three successive targets to output an allophonic label. No improvement has been found so far from traditional classification techniques (i.e. k-nearest neighbors). In a second series of experiments spectral transformations using MLFFA are introduced for the adaptation to new speakers. Compared to linear techniques (multivariate regression and canonical correlation analysis), the MLFFA approach offers some improvement.>

The automatic recognition of continuous speech may use a symbolic representation of the acoustic signal in order to facilitate lexical access. The allophones of the language form a practical set of symbols. A major issue is a reliable localisation of these units in the speech stream and their identification. Localisation is obtained using a robust implementation of temporal decomposition, a technique originally proposed by Atal (1983), for speech coding. Speech is decomposed in terms of overlapping events characterized by both a spectral target and a time-limited interpolation function. An undershot target may be reestimated using neighbours and the associated functions. The possibility of undoing the effects of coarticulation is the major contribution of this work. The identification of these corrected targets is therefore possible with no further contextual rules. The recognition of spelled surnames (letters of the alphabet) is used for evaluation. 76% of correct phones allow 70% of correct letters.>

Localization is realized using a robust implementation of temporal decomposition, a technique originally proposed for speech coding. Speech is decomposed in terms of overlapping events characterized by spectral targets and time-limited interpolation functions. The spectral targets are evaluated iteratively. An undershot target may be reestimated using neighbors and the associated functions. The possibility of undoing the effects of coarticulation is also suggested. The identification of these corrected targets is therefore possible with no further contextual rules. Applications to speech recognition strategies are discussed. The recognition of spelled surnames (letters of the alphabet) is used for evaluation. The correct recognition of 76% of correct phones is the basis of 70% correct recognition letters.>