In the context of speech perception, the cocktail party effect [1, 2] is the ability of the brain to recognize speech in complex and adverse listening conditions where the attended speech is mixed with competing sounds/speech.
This work aims at extracting the speech of a target speaker from single channel audio of several people talking simultaneously. This is an ill-posed problem in that many different hypotheses about what the target speaker says are consistent with the mixture signal. Yet, it can be solved by exploiting some additional information associated to the speaker of interest and/or by leveraging some prior knowledge about speech signal properties (e.g., ). In this work we use face movements of the target speaker as additional information.
This paper (i) proposes the use of face landmark’s movements, extracted using Dlib [6, 7] and (ii) compares different ways of mapping such visual features into time-frequency (T-F) masks, then applied to clean the acoustic mixed-speech spectrogram.
By using Dlib extracted landmarks we relieve our models from the task of learning useful visual features from raw pixels. That aspect is particularly relevant when the training audio-visual datasets are small.
The analysis of landmark-dependent masking strategies is motivated by the fact that speech enhancement mediated by an explicit masking is often more effective than mask-free enhancement .
1.1 Related work
Speech enhancement aims at extracting the voice of a target speaker, while speech separation refers to the problem of separating each sound source in a mixture. Recently proposed audio-only single-channel methods have achieved very promising results [11, 12, 13]
. However the task still remains challenging. Additionally, audio-only systems need separate models in order to associate the estimated separated audio sources to each speaker, while vision easily allow that in a unified model.
Regarding audio-visual speech enhancement and separation methods an extensive review is provided in 
. Here we focus on the deep-learning methods that are most related to the present work.
, where a pre-trained convolutional neural network (CNN) is used to generate a clean spectrogram from silent video. Rather than directly computing a time-frequency (T-F) mask, the mask is computed by thresholding the estimated clean spectrogram. This approach is not very effective since the pre-trained CNN is designed for a different task (video-to-speech synthesis). In  a CNN is trained to directly estimate clean speech from noisy audio and input video. A similar model is used in , where the model jointly generates clean speech and input video in a denoising-autoender architecture.
shows that using information about lip positions can help to improve speech enhancement. The video feature vector is obtained computing pair-wise distances between any mouth landmarks. Similarly to our approach their visual features are not learned on the audio-visual dataset but are provided by a system trained on different dataset. Contrary to our approach, uses position-based features while we use motion features (of the whole face) that in our experiments turned out to be much more effective than positional features.
Although the aforementioned audio-visual methods work well, they have only been evaluated in a speaker-dependent setting. Only the availability of new large and heterogeneous audio-visual datasets has allowed the training of deep neural network-based speaker-independent speech enhancement models [20, 21, 22].
The present work shows that huge audio-visual datasets are not a necessary requirement for speaker-independent audio-visual speech enhancement. Although we have only considered datasets with simple visual scenarios (i.e., the target speaker is always facing the camera), we expect our methods to perform well in more complex scenarios thanks to the robust landmark extraction.
2 Model Architectures
We experimented with the three models shown in Fig. 1. All models receive in input the target speaker’s landmark motion vectors and the power-law compressed spectrogram of the single-channel mixed speech signal. All of them perform some kind of masking operation.
2.1 VL2M model
At each time frame, the video-landmark to mask (VL2M) model (Fig. 1a) estimates a T-F mask from visual features only. Formally, given a video sequence and a target mask sequence , VL2M perform a function , where is the estimated mask.
The training objective for VL2M is a Target Binary Mask (TBM) [23, 24], computed using the spectrogram of the target speaker only. This is motivated by our goal of extracting the speech of a target speaker as much as possible independently of the concurrent speakers, so that, e.g., we do not need to estimate their number.
Given a clean speech spectrogram of a speaker , the TBM is defined by comparing, at each frequency bin , the target speaker value vs. a reference threshold . As in , we use a function of long-term average speech spectrum (LTASS) as reference threshold. This threshold indicates if a T-F unit is generated by the speaker or refers to silence or noise. The process to compute the speaker’s TBM is as follows:
and the standard deviationare computed for all frequency bins of all seen spectrograms in speaker’s data.
The threshold is defined as:
where is a value selected by manual inspection of several spectrogram-TBM pairs.
The threshold is applied to every speaker’s speech spectrogram .
is carried out by a stacked bi-directional Long Short-Term Memory (BLSTM) network. The BLSTM outputs are then forced to lay within the range. Finally the computed TBM and the noisy spectrogram are element-wise multiplied to obtain the estimated clean spectrogram , where .
The model parameters are estimated to minimize the loss:
2.2 VL2M_ref model
VL2M generates T-F masks that are independent of the acoustic context. We may want to refine the masking by including such context. This is what the novel VL2M_ref does (Fig. 1b). The computed TBM and the input spectrogram are the input to a function that outputs an Ideal Amplitude Mask (IAM) (known as FFT-MASK in ). The IAM is defined as:
Note that although IAM generation requires the mixed-speech spectrogram, separate spectrograms for each concurrent speakers are not required.
The target speaker’s spectrogram is reconstructed by multiplying the input spectrogram with the estimated IAM. Values greater than in the IAM are clipped to in order to obtain better numerical stability as suggested in .
The model performs a function that consists of a VL2M component plus three different BLSTMs , and .
receives the VL2M mask as input, and is fed with the noisy spectrogram. Their output are fused in a joint audio-visual representation , where is a linear combination of and :
is the input of the third BLSTM , where lays in the [0,10] range.
The loss function is:
2.3 Audio-Visual concat model
The third model (Fig. 1c) performs early fusion of audio-visual features. This model consists of a single stacked BLSTM that computes the IAM mask from the concatenated . The training loss is the same used to train VL2M_ref. This model can be regarded as a simplification of VL2M_ref, where the VL2M operation is not performed.
3 Experimental setup
Regarding the GRID corpus, for each of the speakers (one had to be discarded) we first randomly selected utterances (out of ). Then, for each utterance, we created different audio-mixed samples. Each audio-mixed sample was created by mixing the chosen utterance with one utterance from a different speaker. That resulted in audio-mixed samples per speaker.
The resulting dataset was split into disjoint sets of // speakers for training/validation/testing respectively.
The TCD-TIMIT corpus consists of speakers (we excluded professionally-trained lipspeakers) and utterances per speaker. The mixed-speech version was created following the same procedure as for GRID, with one difference. Contrary to GRID, TCD-TIMIT utterances have different duration. Thus utterances were mixed only if their duration difference did not exceed seconds. For each utterance pair, we forced the non-target speaker’s utterance to match the duration of the target speaker utterance. If it was longer, the utterance was cut at its end, whereas if it was shorter, silence samples were equally added at its start and end.
The resulting dataset was split into disjoint sets of // speakers for training/validation/testing respectively.
3.2 LSTM training
In all experiments, the models were trained using the Adam optimizer . Early stopping was applied when the error on the validation set did not decrease over
VL2M and Audio-Visual concat had and stacked BLSTM layers respectively. All BLSTMs had units. Hyper-parameters selection was performed by using random search with a limited number of samples, therefore all the reported results may improve through a deeper hyper-parameters validation phase.
VL2M_ref training was performed in steps. We first pre-trained the model using the oracle TBM . Then we substituted the oracle masks with the VL2M component and retrained the model freezing the parameters of the VL2M component.
3.3 Audio pre- and post-processing
The original waveforms were resampled to 16 kHz. Short-Time Fourier Transform (STFT) was computed using FFT size of 512, Hann window of length 25 ms (400 samples), and hop length of 10 ms (160 samples). We then performed power-law compressionwith . Finally we applied per-speaker 0-mean 1-std normalization.
In the post-processing stage, the enhanced waveform generated by the speech enhancement models was reconstructed by applying the inverse STFT to the estimated clean spectrogram and using the phase of the noisy input signal.
3.4 Video pre-processing
Face landmarks were extracted from video using the Dlib  implementation of the face landmark estimator described in . It returns 68 x-y points, for an overall 136 values. We upsampled from 25/29.97 fps (GRID/TCD-TIMIT) to
fps to match the frame rate of the audio spectrogram. Upsampling was carried out through linear interpolation over time.
The final video feature vector was obtained by computing the per-speaker normalized motion vector of the face landmarks by simply subtracting every frame with the previous one. The motion vector of the first frame was set to zero.
|2 Speakers||3 Speakers|
|2 Speakers||3 Speakers|
In order to compare our models to previous works in both speech enhancement and separation, we evaluated the performance of the proposed models using both speech separation and enhancement metrics. Specifically, we measured the capability of separating the target utterance from the concurrent utterance with the source-to-distortion ratio (SDR) [27, 28]. While the quality of estimated target speech was measured with the perceptual PESQ  and ViSQOL  metrics. For PESQ we used the narrow band mode while for ViSQOL we used the wide band mode.
As a very first experiment we compared landmark position vs. landmark motion vectors. It turned out that landmark positions performed poorly, thus all results reported here refer to landmark motion vectors only.
We then carried out some speaker-dependent experiments to compare our models to previous studies as there are no reported results of speaker-independent systems trained and tested on GRID and TCD-TIMIT. Table 1 reports the test-set evaluation of speaker-dependent models on the GRID corpus with landmark motion vectors. Results are comparable with previous state-of-the-art studies in an almost identical setting [15, 17].
Table 2 and 3 show speaker-independent test-set results on the GRID and TCD-TIMIT datasets respectively. V2ML performs significantly worse than the other two models indicating that a successful mask generation has to depend on the acoustic context.
V2ML_ref and AV concat exhibit very similar results, and most importantly, their performance in the speaker-independent setting is comparable to that in the speaker-dependent setting.
Additionally, in order to assess the importance of masking we created a model where a stacked BLSTM directly reconstructs the target speaker spectrogram from without going through any mask operation. During training, we observed a very unstable behavior of the loss function and a SDR value just above on the GRID test-set. Finally, we evaluated the systems in a more challenging testing condition where the target utterance was mixed with utterances from competing speakers. Despite the model was trained with mixtures of two speakers, the decrease of performance was not dramatic.
Code and some testing examples of our models are available at https://goo.gl/3h1NgE.
This paper proposes the use of face landmark motion vectors for audio-visual speech enhancement in a single-channel multi-talker scenario. Different models are tested where landmark motion vectors are used to generate time-frequency (T-F) masks that extract the target speaker’s spectrogram from the acoustic mixed-speech spectrogram.
To the best of our knowledge, some of the proposed models are the first models trained and evaluated on the limited size GRID and TCD-TIMIT datasets that accomplish speaker-independent speech enhancement in the multi-talker setting, with a quality of enhancement comparable to that achieved in a speaker-dependent setting.
-  E. Colin Cherry, “Some experiments on the recognition of speech, with one and with two ears,” The Journal of the Acoustical Society of America, vol. 25, no. 5, pp. 975–979, 1953.
-  Josh H McDermott, “The cocktail party problem,” Current Biology, vol. 19, no. 22, pp. R1024–R1027, 2009.
-  Elana Zion Golumbic, Gregory B. Cogan, Charles E. Schroeder, and David Poeppel, “Visual input enhances selective speech envelope tracking in auditory cortex at a “cocktail party”,” Journal of Neuroscience, vol. 33, no. 4, pp. 1417–1426, 2013.
-  Wei Ji Ma, Xiang Zhou, Lars A. Ross, John J. Foxe, and Lucas C. Parra, “Lip-reading aids word recognition most in moderate noise: A bayesian explanation using high-dimensional feature space,” PLOS ONE, vol. 4, no. 3, pp. 1–14, 03 2009.
-  Albert S Bregman, Auditory scene analysis: The perceptual organization of sound, MIT press, 1994.
-  Vahid Kazemi and Josephine Sullivan, “One millisecond face alignment with an ensemble of regression trees,” in , June 2014.
Davis E. King,
“Dlib-ml: A machine learning toolkit,”Journal of Machine Learning Research, vol. 10, pp. 1755–1758, 2009.
-  Yuxuan Wang, Arun Narayanan, and DeLiang Wang, “On Training Targets for Supervised Speech Separation,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 22, no. 12, pp. 1849–1858, Dec. 2014.
-  Martin Cooke, Jon Barker, Stuart Cunningham, and Xu Shao, “An audio-visual corpus for speech perception and automatic speech recognition,” The Journal of the Acoustical Society of America, vol. 120, no. 5, pp. 2421–2424, Nov. 2006.
-  Naomi Harte and Eoin Gillen, “TCD-TIMIT: An Audio-Visual Corpus of Continuous Speech,” IEEE Transactions on Multimedia, vol. 17, no. 5, pp. 603–615, May 2015.
-  Z. Chen, Y. Luo, and N. Mesgarani, “Deep attractor network for single-microphone speaker separation,” in 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), March 2017, pp. 246–250.
-  Yusuf Isik, Jonathan Le Roux, Zhuo Chen, Shinji Watanabe, and John R. Hershey, “Single-channel multi-speaker separation using deep clustering,” in Interspeech, 2016.
Morten Kolbaek, Dong Yu, Zheng-Hua Tan, Jesper Jensen, Morten Kolbaek, Dong Yu,
Zheng-Hua Tan, and Jesper Jensen,
“Multitalker speech separation with utterance-level permutation invariant training of deep recurrent neural networks,”IEEE/ACM Trans. Audio, Speech and Lang. Proc., vol. 25, no. 10, pp. 1901–1913, Oct. 2017.
-  Bertrand Rivet, Wenwu Wang, Syed Mohsen Naqvi, and Jonathon Chambers, “Audiovisual Speech Source Separation: An overview of key methodologies,” IEEE Signal Processing Magazine, vol. 31, no. 3, pp. 125–134, May 2014.
-  Aviv Gabbay, Ariel Ephrat, Tavi Halperin, and Shmuel Peleg, “Seeing through noise: Visually driven speaker separation and enhancement,” in ICASSP. 2018, pp. 3051–3055, IEEE.
-  Ariel Ephrat, Tavi Halperin, and Shmuel Peleg, “Improved speech reconstruction from silent video,” ICCV 2017 Workshop on Computer Vision for Audio-Visual Media, 2017.
-  Aviv Gabbay, Asaph Shamir, and Shmuel Peleg, “Visual speech enhancement,” in Interspeech. 2018, pp. 1170–1174, ISCA.
-  Jen-Cheng Hou, Syu-Siang Wang, Ying-Hui Lai, Yu Tsao, Hsiu-Wen Chang, and Hsin-Min Wang, “Audio-Visual Speech Enhancement Using Multimodal Deep Convolutional Neural Networks,” IEEE Transactions on Emerging Topics in Computational Intelligence, vol. 2, no. 2, pp. 117–128, Apr. 2018.
-  Jen-Cheng Hou, Syu-Siang Wang, Ying-Hui Lai, Jen-Chun Lin, Yu Tsao, Hsiu-Wen Chang, and Hsin-Min Wang, “Audio-visual speech enhancement using deep neural networks,” in 2016 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA), Jeju, South Korea, Dec. 2016, pp. 1–6, IEEE.
-  Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T. Freeman, and Michael Rubinstein, “Looking to Listen at the Cocktail Party: A Speaker-Independent Audio-Visual Model for Speech Separation,” ACM Transactions on Graphics, vol. 37, no. 4, pp. 1–11, July 2018, arXiv: 1804.03619.
-  T. Afouras, J. S. Chung, and A. Zisserman, “The conversation: Deep audio-visual speech enhancement,” in Interspeech, 2018.
-  Andrew Owens and Alexei A Efros, “Audio-visual scene analysis with self-supervised multisensory features,” European Conference on Computer Vision (ECCV), 2018.
-  Michael C. Anzalone, Lauren Calandruccio, Karen A. Doherty, and Laurel H. Carney, “Determination of the potential benefit of time-frequency gain manipulation,” Ear Hear, vol. 27, no. 5, pp. 480–492, Oct 2006, 16957499[pmid].
-  Ulrik Kjems, Jesper B. Boldt, Michael S. Pedersen, Thomas Lunner, and DeLiang Wang, “Role of mask pattern in intelligibility of ideal binary-masked noisy speech,” The Journal of the Acoustical Society of America, vol. 126, no. 3, pp. 1415–1426, 2009.
-  A. Graves, A. Mohamed, and G. Hinton, “Speech recognition with deep recurrent neural networks,” in 2013 IEEE International Conference on Acoustics, Speech and Signal Processing, May 2013, pp. 6645–6649.
-  Diederik P Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014.
-  E. Vincent, R. Gribonval, and C. Fevotte, “Performance measurement in blind audio source separation,” IEEE Transactions on Audio, Speech and Language Processing, vol. 14, no. 4, pp. 1462–1469, July 2006.
-  Colin Raffel, Brian McFee, Eric J Humphrey, Justin Salamon, Oriol Nieto, Dawen Liang, Daniel PW Ellis, and C Colin Raffel, “mir_eval: A transparent implementation of common mir metrics,” in In Proceedings of the 15th International Society for Music Information Retrieval Conference, ISMIR. Citeseer, 2014.
-  A.W. Rix, J.G. Beerends, M.P. Hollier, and A.P. Hekstra, “Perceptual evaluation of speech quality (PESQ)-a new method for speech quality assessment of telephone networks and codecs,” in 2001 IEEE International Conference on Acoustics, Speech, and Signal Processing. Proceedings (Cat. No.01CH37221), Salt Lake City, UT, USA, 2001, vol. 2, pp. 749–752, IEEE.
-  A. Hines, J. Skoglund, A. Kokaram, and N. Harte, “ViSQOL: The Virtual Speech Quality Objective Listener,” in IWAENC 2012; International Workshop on Acoustic Signal Enhancement, Sept. 2012, pp. 1–4.