This paper introduces the resources necessary to develop and evaluate human assisted lifelong learning speaker diarization systems. It describes the ALLIES corpus and associated protocols, especially designed for diarization of a collection audio recordings across time. This dataset is compared to existing corpora and the performances of three baseline systems, based on x-vectors,i-vectors and VBxHMM, are reported for reference. Those systems are then extended to include an active correction process that efficiently guides a human annotator to improve the automatically generated hypotheses. An open-source simulated human expert is provided to ensure reproducibility of the human assisted correction process and its fair evaluation. An exhaustive evaluation, of the human assisted correction shows the high potential of this approach. The ALLIES corpus, a baseline system including the active correction module and all evaluation tools are made freely available to the scientific community.
Voice activity and overlapped speech detection (respectively VAD and OSD) are key pre-processing tasks for speaker diarization. The final segmentation performance highly relies on the robustness of these sub-tasks. Recent studies have shown VAD and OSD can be trained jointly using a multi-class classification model. However, these works are often restricted to a specific speech domain, lacking information about the generalization capacities of the systems. This paper proposes a complete and new benchmark of different VAD and OSD models, on multiple audio setups (single/multi-channel) and speech domains (e.g. media, meeting...). Our 2/3-class systems, which combine a Temporal Convolutional Network with speech representations adapted to the setup, outperform state-of-the-art results. We show that the joint training of these two tasks offers similar performances in terms of F1-score to two dedicated VAD and OSD systems while reducing the training cost. This unique architecture can also be used for single and multichannel speech processing.
State of the art diarization systems now achieve decent performance but those performances are often not good enough to deploy them without any human supervision. Additionally, most approaches focus on single audio files while many use cases involving multiple recordings with recurrent speakers require the incremental processing of a collection. In this paper, we propose a framework that solicits a human in the loop to correct the clustering by answering simple questions. After defining the nature of the questions for both single file and collection of files, we propose two algorithms to list those questions and associated stopping criteria that are necessary to limit the work load on the human in the loop. Experiments performed on the ALLIES dataset show that a limited interaction with a human expert can lead to considerable improvement of up to 36.5% relative diarization error rate (DER) for single files and 33.29% for a collection.
Our main goal is to study the interactions between speakers according to their gender and role in broadcast media. In this paper, we propose an extensive study of gender and overlap annotations in various speech corpora mainly dedicated to diarisation or transcription tasks. We point out the issue of the heterogeneity of the annotation guidelines for both overlapping speech and gender categories. On top of that, we analyse how the speech content (casual speech, meetings, debate, interviews, etc.) impacts the distribution of overlapping speech segments. On a small dataset of 93 recordings from LCP French channel, we intend to characterise the interactions between speakers according to their gender. Finally, we propose a method which aims to highlight active speech areas in terms of interactions between speakers. Such a visualisation tool could improve the efficiency of qualitative studies conducted by researchers in human sciences.
This article focuses on overlapped speech and gender detection in order to study interactions between women and men in French audiovisual media (Gender Equality Monitoring project). In this application context, we need to automatically segment the speech signal according to speakers gender, and to identify when at least two speakers speak at the same time. We propose to use WavLM model which has the advantage of being pre-trained on a huge amount of speech data, to build an overlapped speech detection (OSD) and a gender detection (GD) systems. In this study, we use two different corpora. The DIHARD III corpus which is well adapted for the OSD task but lack gender information. The ALLIES corpus fits with the project application context. Our best OSD system is a Temporal Convolutional Network (TCN) with WavLM pre-trained features as input, which reaches a new state-of-the-art F1-score performance on DIHARD. A neural GD is trained with WavLM inputs on a gender balanced subset of the French broadcast news ALLIES data, and obtains an accuracy of 97.9%. This work opens new perspectives for human science researchers regarding the differences of representation between women and men in French media.
à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d'enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
End-to-end neural architectures for spoken language understanding In this paper, we focus on end-to-end architectures designed to tackle spoken language understanding problems. We propose encoder-decoder architectures with an attention mechanism that focuses on relevant contextual acoustic features to hypothesize semantic contents. A first architecture has been built in order to extract pronounced words and concepts from speech. Tested on the MEDIA dataset, it obtains good results, and combined with a language model, it lowers the error by 2.8 points from the state-of-the-art results with end-to-end systems. With this same architecture, we propose a new configuration allowing to predict both concepts and their values. Lastly, a new architecture is proposed, composed of multiple chained decoders for a single encoder. With this architecture, we aim to improve the decoder with both linguistic and acoustic informations. MOTS-CL  ES : compr ehension de la parole, r eseaux de neurones, m ecanisme d’attention, s equence vers s equence, transfert d’apprentissage.
This papers aims at improving spoken language modeling (LM) using very large amount of automatically transcribed speech. We leverage the INA (French National Audiovisual Institute(1)) collection and obtain 19GB of text after applying ASR on 350,000 hours of diverse TV shows. From this, spoken language models are trained either by fine-tuning an existing LM (FlauBERT(2)) or through training a LM from scratch. The new models (FlauBERT-Oral) are shared with the community(3) and are evaluated not only in terms of word prediction accuracy but also for two downstream tasks: classification of TV shows and syntactic parsing of speech. Experimental results show that FlauBERT-Oral is better than its initial FlauBERT version demonstrating that, despite its inherent noisy nature, ASR-Generated text can be useful to improve spoken language modeling.
Word embedding methods allow to represent words as vectors in a space that is structured using word co-occurrences so that words with close meanings are close in this space. These vectors are then provided as input to automatic systems to solve natural language processing problems. Because interpretability is a necessary condition to trusting such systems, interpretability of embedding spaces, the first link in the chain is an important issue. In this paper, we thus evaluate the interpretability of vectors extracted with two approaches: SPINE, a k-sparse auto-encoder, and SINr, a graph-based method. This evaluation is based on a Word Intrusion Task with human annotators. It is operated using a large French corpus, and is thus, as far as we know, the first large-scale experiment regarding word embedding interpretability on this language. Furthermore, contrary to the approaches adopted in the literature where the evaluation is performed on a small sample of frequent words, we consider a more realistic use-case where most of the vocabulary is kept for the evaluation. This allows to show how difficult this task is, even though SPINE and SINr show some promising results. In particular, SINr results are obtained with a very low amount of computation compared to SPINE, while being similarly interpretable.
State of the art diarization systems now achieve decent performance but those performances are often not good enough to deploy them without any human supervision. In this paper we propose a framework that solicits a human in the loop to correct the clustering by answering simple questions. After defin-ing the nature of the questions, we propose an algorithm to list those questions and two stopping criteria that are necessary to limit the work load on the human in the loop. Experiments performed on the ALLIES dataset show that a limited interaction with a human expert can lead to considerable improvement of up to 36.5% relative diarization error rate (DER) compared to a strong baseline.
Diarization consists in the segmentation of speech signals and the clustering of homogeneous speaker segments. State-of-the-art systems typically operate upon speaker embeddings, such as i-vectors or neural x-vectors, extracted from mel cepstral coefficients (MFCCs) or spectrograms. The recent SincNet architecture extracts x-vectors directly from raw speech signals. The work reported in this paper compares the performance of different embeddings extracted from MFCCs or the raw signal for speaker diarization and broadcast media treated with compression and sub-sampling, operations which typically degrade performance. Experiments are performed with the new ALLIES database that was designed to complement existing, publicly available French corpora of broadcast radio and TV shows. Results show that, in adverse conditions, with compression and sampling mismatch, SincNet x-vectors outperform i-vectors and x-vectors by relative DERs of 43% and 73% respectively. Additionally we found that SincNet x-vectors are not the absolute best embeddings but are more robust to data mismatch than others.
The ANTRACT project is a cross-disciplinary apparatus dedicated to the analysis of the French newsreel company Les Actualites Francaises (1945-1969) and its film productions. Founded during the liberation of France, this state-owned company filmed more than 20,000 news reports shown in French cinemas and throughout the world over its 24 years of activity. The project brings together research organizations with a dual historical and technological perspective. ANTRACT's goal is to study the production process, the film content, the way historical events are represented and the audience reception of Les Actualites Francaises newsreels using innovative AI-based data processing tools developed by partners specialized in image, audio, and text analysis. This article focuses on the data processing apparatus and tools of the project. Automatic content analysis is used to select data, to segment video units and typescript images, and to align them with their archival description. Automatic speech recognition provides a textual representation and natural language processing can extract named entities from the voice-over recording; automatic visual analysis is applied to detect and recognize faces of well-known characters in videos. These multifaceted data can then be queried and explored with the TXM text-mining platform. The results of these automatic analysis processes are feeding the Okapi platform, a client-server software that integrates documentation, information retrieval, and hypermedia capabilities within a single environment based on the Semantic Web standards. The complete corpus of Les Actualites Francaises, enriched with data and metadata, will be made available to the scientific community by the end of the project.
In this paper, we propose a novel end-to-end sequence-to-sequence spoken language understanding model using an attention mechanism. It reliably selects contextual acoustic features in order to hypothesize semantic contents. An initial architecture capable of extracting all pronounced words and concepts from acoustic spans is designed and tested. With a shallow fusion language model, this system reaches a 13.6 concept error rate (CER) and an 18.5 concept value error rate (CVER) on the French MEDIA corpus, achieving an absolute 2.8 points reduction compared to the state-of-the-art. Then, an original model is proposed for hypothesizing concepts and their values. This transduction reaches a 15.4 CER and a 21.6 CVER without any new type of context.
While graph embedding aims at learning low-dimensional representations of nodes encompassing the graph topology, word embedding focus on learning word vectors that encode semantic properties of the vocabulary. The first finds applications on tasks such as link prediction and node classification while the latter is systematically considered in natural language processing. Most of the time, graph and word embeddings are considered on their own as distinct tasks. However, word co-occurrence matrices, widely used to extract word embeddings, can be seen as graphs. Furthermore, most network embedding techniques rely either on a word embedding methodology (Word2vec) or on matrix factorization, also widely used for word embedding. These methods are usually computationally expensive, parameter dependant and the dimensions of the embedding space are not interpretable. To circumvent these issues, we introduce the Lower Dimension Bipartite Graphs Framework (LDBGF) which takes advantage of the fact that all graphs can be described as bipartite graphs, even in the case of textual data. This underlying bipartite structure may be explicit, like in coauthor networks. However, with LDBGF, we focus on uncovering latent bipartite structures, lying for instance in social or word co-occurrence networks, and especially such structures providing conciser and interpretable representations of the graph at hand. We further propose SINr, an efficient implementation of the LDBGF approach that extracts Sparse Interpretable Node Representations using community structure to approximate the underlying bipartite structure. In the case of graph embedding, our near-linear time method is the fastest of our benchmark, parameter-free and provides state-of-the-art results on the classical link prediction task. We also show that low-dimensional vectors can be derived from SINr using singular value decomposition. In the case of word embedding, our approach proves to be very efficient considering the classical similarity evaluation.
We developed a human assisted speaker diarization platform that enables a human annotator to correct the output of any speaker diarization system by providing a graphical view of the diarization segmentation and clustering steps while guiding the human annotator to optimize the correction process and easily improve the resulting diarization.
In this paper, we present a framework to evaluate the human corrections of a speaker diarization system. We propose four elementary actions to correct the diarization (“ Create a boundary ”, “ Delete a boundary ”, “ Create a speaker label ” and “ Change the speaker label ”) and we propose an automaton to simulate the correction sequence. A metric is described to evaluate the correction cost. The framework is evaluated using French broadcast news drawn from the following campaigns: REPERE, ESTER and ETAPE.
Aujourd’hui les systemes intelligents obtiennent d’excellentes performances dans de nombreux domaines lorsqu’ils sont entraines par des experts en apprentissage automatique. Lorsque ces systemes sont mis en production, leurs performances se degradent au cours du temps du fait de l’evolution de leur environnement reel. Une adaptation de leur modele par des experts en apprentissage automatique est possible mais tres couteuse alors que les societes utilisant ces systemes disposent d’experts du domaine qui pourraient accompagner ces systemes dans un apprentissage tout au long de la vie. Dans cet article nous proposons un cadre d’evaluation generique pour des systemes apprenant tout au long de la vie (SATLV). Nous proposons d’evaluer l’apprentissage assiste par l’humain (actif ou interactif) et l’apprentissage au cours du temps.
Current intelligent systems need the expensive support of machine learning experts to sustain their performance level when used on a daily basis. To reduce this cost, i.e. remaining free from any machine learning expert, it is reasonable to implement lifelong (or continuous) learning intelligent systems that will continuously adapt their model when facing changing execution conditions. In this work, the systems are allowed to refer to human domain experts who can provide the system with relevant knowledge about the task. Nowadays, the fast growth of lifelong learning systems development rises the question of their evaluation. In this article we propose a generic evaluation methodology for the specific case of lifelong learning systems. Two steps will be considered. First, the evaluation of human-assisted learning (including active and/or interactive learning) outside the context of lifelong learning. Second, the system evaluation across time, with propositions of how a lifelong learning intelligent system should be evaluated when including human assisted learning or not.
Jean-Marc Odobez合作论文数Perception & Activity Understanding Group, Idiap Research Institute;École Polytechnique Fédérale de Lausanne9