An ideal quantized mask to increase intelligibility and quality of speech in noise

Abstract

Time-frequency (T-F) masks represent powerful tools to increase the intelligibility of speech in background noise. Translational relevance is provided by their accurate estimation based only on the signal-plus-noise mixture, using deep learning or other machine-learning techniques. In the current study, a technique is designed to capture the benefits of existing techniques. In the ideal quantized mask (IQM), speech and noise are partitioned into T-F units, and each unit receives one of N attenuations according to its signal-to-noise ratio. It was found that as few as four to eight attenuation steps (IQM 4 , IQM 8) improved intelligibility over the ideal binary mask (IBM, having two attenuation steps), and equaled the intelligibility resulting from the ideal ratio mask (IRM, having a theoretically infinite number of steps). Sound-quality ratings and rankings of noisy speech processed by the IQM 4 and IQM 8 were also superior to that processed by the IBM and equaled or exceeded that processed by the IRM. It is concluded that the intelligibility and sound-quality advantages of infinite attenuation resolution can be captured by an IQM having only a very small number of steps. Further, the classification-based nature of the IQM might provide algorithmic advantages over the regressionbased IRM during machine estimation. V

Determining the energetic and informational components of speech-on-speech masking in listeners with sensorineural hearing loss

“…ITFS processing usually assumes a priori knowledge of the target and masker signals so that the energy relations in each time-frequency (T-F) unit may be specified exactly and, in that sense, is fundamentally different than normal human perception. The use of ITFS to separate EM from IM is based on the premise that listeners extracting speech in a mixture of sounds likely rely only (or predominantly) on the subset of T-F units in which the energy of the target source relative to the energy of the masking source(s) (e.g., Cooke, 2006;Li and Loizou, 2007;Healey and Vasko, 2018) exceeds a specified value (termed the level criterion, or LC). The T-F units in which masker energy dominates target energy logically fall under the definition of EM because of the assumption that the neural response in a small T-F unit would be driven by the properties (i.e., amplitude and timing) of the higher-energy source.…”

Towards An ASR Approach Using Acoustic and Language Models for Speech Enhancement

“…To further investigate the effectiveness of quantized masking approaches, we use the Chimera++ model to estimate the IQM [13], where the models are Chi++ IQM2 , Chi++ IQM3 , Chi++ IQM4 , and Chi++ IQM8 which predicts enhanced speech using IQM2, IQM3, IQM4, and IQM8 respectively. Here X in IQMX refers to the number of attenuation levels in the mask (see [13]). Additionally, we compare with a Chimera++ network that predicts quantized speech Chi++ quant (proposed without the QSM).…”

“…In [12], separate discrete T-F masks for magnitude and phase responses are estimated using softmax activations, where recurrent networks are used to capture temporal correlations. The ideal quantized mask (IQM) has also recently been proposed [13]. It shows that quantization of the IRM, by coding each T-F IRM value into one of a number of quantized bins, is a reasonable representation of the IRM as assessed by human listeners.…”

Avaliação da Predição Objetiva da Inteligibilidade de Sinais Reverberantes e Ruidosos com Uso de Máscaras Acústicas

“…3) IQM: A máscara IQM [8] utiliza IRM(k, q) como base para a obtenção de N possíveis valores de IQM(k, q). Isto é feito a partir de uma seleção de pontos da máscara IRM de acordo com:…”

“…A máscara IRM (Ideal Ratio Mask) [7] define valores fracionários de ganho em cada quadro TF de acordo com a razão entre a energia do sinal limpo e a do sinal corrompido. Baseada na IRM, a máscara ideal IQM (Ideal Quantized Mask) utiliza valores discretos de níveis de atenuação [8]. Entre as máscaras cegas, a BRM (Binary Reverberant Mask) apresentou melhora de inteligibilidade para ambientes com reverberação em testes subjetivos [9].…”