Descrizione del progetto
Una tecnologia di riconoscimento automatico del parlante ottimizzata
Il riconoscimento vocale si configura come una funzione primaria per una vasta gamma di applicazioni. Il crescente sviluppo di tecniche per l’utilizzo e l’analisi di dati offre soluzioni intese a migliorare in modo continuativo l’industria dell’elaborazione vocale. Il progetto ETE SPEAKER, finanziato dall’UE, si propone di sviluppare uno strumento innovativo basato sul riconoscimento automatico del parlante che isola le informazioni necessarie a determinare l’identità di chi sta parlando durante una registrazione vocale. ETE SPEAKER concentrerà l’attenzione su uno studio e utilizzo dettagliato delle potenzialità insite nelle reti neurali profonde allo scopo di distinguere le informazioni relative al parlante dal resto della variabilità dei disturbi sonori. Il suo obiettivo è quello di introdurre una soluzione di riconoscimento automatico del parlante completa che si allinei agli ultimi standard in materia di valutazione del riconoscimento del parlante.
Obiettivo
This project focuses on automatic speaker recognition (SID), the task of determining the identity of the speaker in a speech recording. Disentangling the speaker specific information from the rest of nuisance variability requires complex models. Deep neural networks (DNNs) have recently showed their potential for this, as the popular x-vector learnt by a DNN.
Here, we aim for end-to-end SID where the system is optimized as a whole for the target task. Despite several attempts in this line of research, many aspects still remain unexplored or not explored thoroughly.
We also propose to explore recurrent approaches, suitable for dealing with temporal signals, as well as different pooling methods to obtain a fixed-length representation from a variable length input sequence of speech features.
Next, we want to explore different flavors of attention mechanisms, which make the DNN to focus on relevant parts of the input, providing a way to quantify how much evidence has been collected about the speaker identity and the uncertainty of the obtained representation, which is a critical issue when making (Bayesian) decisions in SID.
Finally, some other approaches such as using the raw signal (instead of features) or other advances that might arise will be also explored for SID and related tasks.
To achieve our goals, we will start from theory, implement the proposed approaches and test on public SID benchmarks such as NIST SREs. The outcomes are intended to benefit both scientific community and speech processing industry.
The applicant Dr. Alicia Lozano-Diez is an excellent female researcher, who has done her Ph.D. at Audias (Universidad Autonoma de Madrid, Spain), a respected research lab. The host group Speech@FIT from Brno University of Technology (Czechia) has a top-class track on speech processing research. Thus, we expect the combination of both the researcher and the host to boost the researcher career and benefit the host group (and its industrial European partners).
Campo scientifico
Programma(i)
Argomento(i)
Meccanismo di finanziamento
MSCA-IF-EF-ST - Standard EFCoordinatore
601 90 Brno Stred
Cechia