A Fully Convolutional Neural Network Approach to End-to-End Speech Enhancement

A FULL Y CONV OLUTIONAL NEURAL NETW ORK APPRO A CH TO END-T O-END SPEECH ENHANCEMENT F r ank Longueira, Sam K eene The Cooper Union franklongueira@gmail.com, keene@cooper .edu ABSTRA CT This paper will describe a no vel approach to the cocktail party problem that relies on a fully con volutional neural network (FCN) architecture. The FCN takes noisy audio data as input and performs nonlinear , filtering operations to produce clean audio data of the target speech at the output. Our method learns a model for one specific speaker , and is then able to extract that speakers v oice from babble background noise. Results from experimentation indicate the ability to gen- eralize to ne w speakers and robustness to ne w noise en viron- ments of varying signal-to-noise ratios. A potential application of this method would be for use in hearing aids. A pre-trained model could be quickly fine tuned for an indi viduals f amily members and close friends, and deplo yed onto a hearing aid to assist listeners in noisy en vironments. Index T erms — speech enhancement, end-to-end, con vo- lutional neural networks, signal processing, hearing aid 1. INTR ODUCTION One of the largest issues facing hearing impaired indi viduals in their day-to-day lives is accurately recognizing speech in the presence of background noise [ 1 ]. While modern hearing aids do a good job of amplifying sound, the y do not do enough to increase speech quality and intelligibility . This is not a problem in quiet en vironments, but a standard hearing aid that simply amplifies audio will fail to provide the user with a signal they can easily understand when the user is in a noisy en vironment [ 2 ]. The problem of speech intelligibility is even more dif ficult if the background noise is also speech, such as in a bar or restaurant with many patrons. While people without hearing impairments usually have no trouble focusing on a single speaker out of multiple, it is a much more difficult task for people with a hearing impairment [ 3 ]. The problem of picking out one person’ s speech in an en- vironment with many speak ers was dubbed the cocktail party problem [ 4 ]. The paper asserts that humans are normally capa- ble of separating multiple speakers and focusing on a single one. Howe ver , hearing impaired indi viduals may ha ve issues when it comes to performing this same task. A solution to the cocktail party problem would be an algorithm that a computer can employ in real-time to enhance the speech corrupted by babble (background noise from other speakers). T radition- ally , the cocktail party problem has been approached using sev eral different techniques, such as using microphone arrays, monaural algorithms in volving signal processing techniques, and Computational Auditory Scene Analysis (CASA) [1]. Deep learning approaches to the cocktail party problem and speech enhancement in general tend to take noisy spectro- grams as input and transform them to clean spectrograms. The use of deep conv olutional neural networks and deep denoising autoencoders on spectrograms ha ve prov en to be powerful techniques in practice [ 5 ]. One drawback to the use of spec- trograms as input is the computation of spectrograms tends to be high since the short-time Fourier transform has to be applied to the ra w audio data. This prior computation before inputting into the network requires time and hence increases the dif ficulty of use in real-time applications. In addition, phase information of the input speech tends to be lost in man y of these approaches since only the magnitude spectrum is used. This can cause degradation in quality at the output of the system [6]. More recent deep learning approaches ha ve considered an end-to-end approach to speech enhancement that requires no feature extraction [ 7 ] - [ 9 ]. The noisy time-domain au- dio signal is used as the input to a neural network and a fil- tered time-domain audio signal is obtained at the output. This methodology remov es the need for a prior STFT computation and retains phase information at the output. This recent push in the deep learning community to wards end-to-end speech enhancement systems is one of the moti vations for this paper’ s approach. The other large moti vation comes from two papers dealing with the study of CNNs on raw audio data. In the first paper [ 10 ], the authors make a strong case for the lack of need for fully connected layers in a neural netw ork that processes raw audio data at the input. Instead, they recommend the use of con volutional layers in order to maintain local correlations in the signal as it passes through the netw ork. In addition, a fully con volutional netw ork (i.e. a CNN with no fully connected layers) will generally hav e much fe wer parameters than a cor- respondingly similar network that includes fully connected layers. This reduced model comple xity is especially important for real-time application of the speech enhancement algorithm. In the second paper [ 11 ], the authors provide insight into the inner workings of con volutional layers applied to raw audio data. They make a strong case for the lack of need for pooling layers and emphasize the con volution theorem: x ∗ h = F − 1 { F { x } · F { h }} (1) where x can be vie wed as the input audio signal, h is a learned filter , and F is the F ourier transform operator . The con volution theorem allo ws an FCN to be vie wed as a large, nonlinear filter bank. By maintaining the size of the raw audio input vector throughout intermediary computations, each filter’ s output can be vie wed as providing a nonlinear filtered representation of the input vector . As the depth of the FCN increases, a larger number of nonlinear filtered representations is achieved. At the final filtering layer , these representations are combined in a matter that rids the input signal of the background noise rep- resentations and only keeps the tar get speech representations. W ith the motiv ation of the approach in mind, the ne xt section will provide specific details of the system design. 2. SYSTEM DESIGN The first step in designing the speech enhancement system is gathering data for training and validation purposes. An openly av ailable audiobook (narrated by a speaker named Pamela) found online serves as the target speech for designing the system [ 12 ]. In addition, babble noise audio clips were found online to serve as background noise when additiv ely combined to Pamela’ s speech [ 13 ] - [ 14 ]. All of these audio clips were do wnsampled to 16 kHz and to ha ve only one audio channel (taking the element-wise av erage of the two channels if necessary). T able 1 concisely describes this data and how it is split for training and validation. T arget Speech Babble Noise SNR T ime (Min:Sec) T raining Set Chapter 1 Bar Noise 5 dB 35:37 V alidation Set Chapter 2 Cafe Noise 5 dB 5:04 T able 1 . Data collection and splitting for system design pur- poses. T ar get speech refers to Pamela’ s narration of Chapters 1 - 2 in [ 12 ]. Babble noise refers to two dif ferent en vironments found online [ 13 ] - [ 14 ]. Each set of target speech is additively combined with its corresponding set of babble noise at an SNR of 5 dB. It is important to note that the system is being designed around a single speaker (i.e. P amela) and a single SNR of 5 dB. The reason for doing this is to first find a few reasonable FCN architectures for the task of denoising Pamela’ s speech that has been corrupted by babble noise at an SNR of 5 dB. After choosing a subset of FCN architectures, further e xploration will be done for denoising P amela’ s speech at SNRs of 0 dB and -5 dB in order to choose one FCN architecture for the system. Once a single FCN architecture is selected and fixed, the next step will inv olve exploring generalization to a new speaker and the system’ s robustness to different signal-to-noise ratios. Having decided on training and v alidation data, designing the system relies on three main things: (1) a methodology for pre-processing the ra w audio data for input into the FCN, (2) fixing an FCN architecture, (3) a methodology for post- processing the ra w audio data that is output by the FCN. T o begin, let’ s discuss (1). First, tar get speech is additiv ely com- bined with its corresponding babble noise. Next, based on stationarity assumptions for speech, the noisy audio data is split into 20 ms frames in which consecuti ve frames ov erlap by 50%. Each noisy frame is to be multiplied by a corre- sponding Hanning windo w of equal length. T o complete the pre-processing methodology , for each noisy frame the mean of the entire training target speech is element-wise subtracted and standard deviation of the entire training target speech is element-wise di vided. Ha ving fixed the methodology for pre-processing, the methodology for post-processing, (3), can also be fixed if it is assumed the FCN outputs a preprocessed filtered 20 ms frame of the preprocessed noisy 20 ms input frame. Given an output frame from the FCN, it is element- wise multiplied by the standard deviation of the entire training target speech and the mean of the entire training target speech is added element-wise. Next, the output from the FCN for the next frame (k eeping in mind that these two consecuti ve frames o verlap by 50%) is obtained and the same thing is done. Finally , the overlap-add method of reconstruction is applied to undo the Hanning window that w as applied to both ov erlapping frames. This results in having reconstructed 30 ms of filtered raw audio data that is ready for playback. This post-processing methodology can be iteratively done for an ar- bitrary amount of noisy input audio data. The most important part of the system design is (2), fixing an FCN architecture. Since a fully con volutional neural netw ork contains only con volutional layers, the only things to be determined are ho w deep the network needs to be and the details of each layer (i.e. number of filters, kernel size, etc.). The output of the network is to be a one dimensional vector of the same length as the 20 ms input vector , so two things can be immediately concluded upon. The first conclusion is to use “same” padding in all layers to ensure the temporal length of the input v ector remains the same throughout intermediary computations and therefore at the output. This allo ws the FCN to be viewed as a nonlinear , filter bank. The second conclusion is the output layer will be a con volutional layer with one filter and no activ ation function. This output layer is suitable for reconstructing audio data from its nonlinear representations (i.e. the output layer is able to match the range in which audio data exists) and allo ws for an output that is one-dimensional. Next, the kernel size for all filters in the network will initially be fixed at 5 ms in length, i.e. 25% of the input size. This can be tuned later on via the validation set. In addition, a dilation factor will not be used in any conv olutional layer in order to better preserve local correlations. The structure of each hidden layer will be the following: con volution operation, batch normalization, and ReLU (or PReLU) acti vation. Batch normalization between the conv olution operation and activ ation function tends to improv e training time and generalization performance [ 17 ]. Also, the ReLU (or PReLU) acti vation tends to w ork well in general CNN practice [ 15 ]. W ith all of this covered, the only thing left to determine is how man y hidden layers are necessary and how man y filters per hidden layer . These hyperparameters will be determined by training dif ferent FCN architectures and comparing MSE performance on the v alidation set. All models to follow are supervised trained with Adam SGD to minimize MSE [ 16 ]. In addition, all training employs early stopping that terminates after 20 epochs of no impro vement and returns the parameters of the model with best v alidation loss during training. T o begin, an architecture with one hidden layer is trained to find the number of filters needed in a layer . The number of filters is slowly increased and with each ne w number of filters a model is trained and validation loss computed. This process provides an understanding of ho w much complexity , in terms of number of filters per layer , is needed for this task. Number of Filters T raining MSE V alidation MSE 50 0.0401 0.0541 100 0.0384 0.0512 200 0.0398 0.0519 300 0.0383 0.0504 400 0.0372 0.0496 500 0.0377 0.0490 600 0.0381 0.0504 700 0.0384 0.0503 800 0.0375 0.0507 900 0.0391 0.0497 1000 0.0373 0.0498 T able 2 . This table describes training loss & v alidation loss (MSE) of single hidden layer FCN architectures as the number of filters increases. T able 2 sho ws that increasing the number of filters marginally helps reduce training loss and validation loss. W ith this result, a similar procedure will be follo wed to gain an understanding of ho w the depth of the netw ork impro ves performance. The procedure in volves first fixing the number of filters to be either 50, 100, or 200 filters per layer and then the depth of the network is increased. Fig. 1 . A plot that shows increasing network depth leads to large decreases in validation loss. Each curve was generated by fixing the number of filters per hidden layer (either 50, 100, or 200) and increasing the number of hidden layers in the network. From Figure 1, it can be concluded that increasing depth largely helps decrease v alidation loss as compared to increas- ing the number of filters in a giv en layer . With the kno wledge that between 1 to 200 filters in a layer and a depth of 5 to 6 layers tends to provide good v alidation loss, dif ferent net- work architectures are experimented with within this range. Architectures with dif ferent kernel sizes, acti vation functions, number of filters and re gularization were experimented with with the goal of minimizing validation loss. Over 70 FCN architectures were trained and v alidated in total. From the o ver 70 FCN architectures, the top 13 FCN architectures that pro- vide the lo west v alidation loss are taken to be studied further . The PESQ and WER of the filtered validation set is computed for each of the 13 architectures and compared to the PESQ and WER of the noisy validation set (at an SNR of 5 dB) [ 19 ] - [20]. These results are displayed in T able 3. T able 4 presents the architecture from T able 3 (Model #53) that provided the best combination of high validation PESQ (good speech quality), low validation WER (good intelligi- bility), and sufficient model complexity for learning across dif ferent SNRs. This FCN architecture will be used for testing the speech enhancement system. The next section will present results related to testing the system on ne w audio data in order to measure its ability to generalize on the same speaker . Model # Number of Parameters PESQ WER 25 3,218,501 2.470 29.001% 26 12,837,001 2.445 28.591% 27 1,009,501 2.390 26.402% 30 1,209,751 2.441 31.464% 31 4,819,501 2.490 31.737% 33 2,142,896 2.421 28.728% 38 5,867,728 2.452 29.001% 41 4,682,896 2.422 28.454% 53 2,266,736 2.458 25.718% 64 761,251 2.437 29.275% 69 761,501 2.443 27.633% 70 1,562,101 2.451 29.412% 71 841,251 2.480 27.223% T able 3 . Number of parameters, PESQ, and WER of top 13 FCN architectures in terms of validation loss. The specific details of each architecture are remov ed for brevity . For com- parison, the noisy v alidation set at an SNR of 5 dB has a PESQ of 1.764 and WER of 50.479%. Layer T ype Output Shape Number of Parameters 1-D Con volution (320, 12) 972 Batch Normalization (320, 12) 48 PReLU Activ ation (320, 12) 3,840 1-D Con volution (320, 25) 24,025 Batch Normalization (320, 25) 100 PReLU Activ ation (320, 25) 8,000 1-D Con volution (320, 50) 100,050 Batch Normalization (320, 50) 200 PReLU Activ ation (320, 50) 16,000 1-D Con volution (320, 100) 400,100 Batch Normalization (320, 100) 400 PReLU Activ ation (320, 100) 32,000 1-D Con volution (320, 200) 1,600,200 Batch Normalization (320, 200) 800 PReLU Activ ation (320, 200) 64,000 1-D Con volution (320, 1) 16,001 T able 4 . A layer-by-layer description of the speech enhance- ment system’ s FCN architecture (Model #53 from T able 3). 3. TESTING GENERALIZA TION ON THE SAME SPEAKER The first part of testing in volv es training the speech enhance- ment system on a speak er and testing it on the same speaker , but the speech will hav e nev er been seen by the system nor the babble en vironment. First, the speech enhancement system trained and v alidated from the previous section, i.e. using the model from T able 4 as the FCN architecture, is used for testing. Fiv e minutes from Chapter 3 from [ 12 ] will be used as the test target speech and fiv e minutes of a ne w babble envi- ronment [ 18 ] will be used for the test background noise. T o make sure the testing procedure is clear, consider the following walkthrough of the process. First, using the training setup of the pre vious section, the speech enhancement system is trained using Chapter 1 from [ 12 ] and bar babble noise from [ 13 ] at a specific SNR, say 5 dB. The training process employs early stopping that uses 5 minutes from Chapter 2 from [ 12 ] and 5 minutes of cafe babble noise from [ 14 ] at the same SNR. Next, the trained system is tested on 5 minutes from Chapter 3 in [ 12 ] and 5 minutes of coffee shop babble noise from [ 18 ] at the same SNR (i.e. 5 dB), b ut also at 0 dB and -5 dB to get a measure of the system’ s robustness across SNRs. This process is repeated for an SNR of 0 dB and an SNR of -5 dB. T ables 5 & 6 report the results of this process. T esting SNR 5 dB 0 dB -5 dB T raining SNR 5 dB 2.478 1.917 1.350 0 dB 2.530 2.100 1.461 -5 dB 2.379 2.060 1.482 (Noisy: 1.782) (Noisy: 1.444) (Noisy: 1.321) T able 5 . This table presents PESQ test results across dif ferent SNRs for testing on the same speak er . Each ro w represents the SNR that the speech enhancement system was trained at. Each column represents the SNR of the test set. For reference, the PESQ of the noisy test set is included for each SNR. T esting SNR 5 dB 0 dB -5 dB T raining SNR 5 dB 25.243% 60.888% 94.868% 0 dB 31.900% 58.391% 91.817% -5 dB 52.705% 77.531% 94.452% (Noisy: 43.689%) (Noisy: 86.269%) (Noisy: 97.503%) T able 6 . This table presents WER test results across different SNRs for testing on the same speaker . Each ro w represents the SNR the speech enhancement system was trained at. Each column represents the SNR of the test set. For reference, the WER of the noisy test set is included for each SNR. T ables 5 & 6 provide some interesting insights into the generalizability of the speech enhancement system. The first key insight is that both PESQ and WER on the test set do a good job tracking the results for PESQ and WER seen on the validation set. The other key insight is that the speech enhancement system trained at 0 dB is quite robust across SNRs, sometimes doing better in scenarios one would not e x- pect. With promising results from testing on the same speaker , the second part of testing will study generalizability to a new speaker . 4. TESTING GENERALIZA TION ON A NEW SPEAKER The second part of testing will employ the same methodology as the first part of testing but the target speech will come from a new speak er . Audio data is acquired for a new speaker (specifically another female speaker by t he name of T ricia) via another audiobook [ 21 ]. First, performance will be measured when the speech enhancement system is trained only on the speaker Pamela (as has been done up to this point) and tested on the speaker T ricia. All trained models (i.e. at each specific SNR) from the first part of testing are used again in the second part of testing. These trained models filter 5 minutes of speech from Chapter 1 of [ 21 ] corrupted by the same cof fee shop babble noise [ 18 ] at SNRs of -5, 0, and 5 dB. T ables 7 & 8 report the results of this process. T esting SNR 5 dB 0 dB -5 dB T raining SNR 5 dB 2.215 1.781 1.291 0 dB 2.171 1.839 1.382 -5 dB 2.229 1.876 1.437 (Noisy: 1.874) (Noisy: 1.471) (Noisy: 1.182) T able 7 . This table presents PESQ test results across dif ferent SNRs for training on one speaker and testing on a new speaker . T esting SNR 5 dB 0 dB -5 dB T raining SNR 5 dB 25.134% 54.545% 89.037% 0 dB 29.412% 50.401% 84.358% -5 dB 55.615% 67.380% 89.305% (Noisy: 33.556%) (Noisy: 72.326%) (Noisy: 94.652%) T able 8 . This table presents WER test results across different SNRs for training on one speaker and testing on a new speaker . When comparing T ables 5 & 6 with T ables 7 & 8, it is no- ticed that performance on the new speaker is good but does not quite track the performance of a system trained on a speaker and then tested on that same speaker . It is h ypothesized that fine-tuning the parameters of the trained network with a few minutes of data from the new speaker should impro ve per- formance and more closely track performance on the same speaker . Therefore, an additional, disjoint 5 minutes of speech from Chapter 1 of [ 21 ] corrupted by 5 minutes of cafe babble noise from [ 14 ] at a gi ven SNR is used to fine-tune the already trained model (i.e. trained on the speak er Pamela) by doing 5 epochs of gradient descent. The resulting trained model is then used to ag ain filter 5 minutes of speech from Chapter 1 of [ 21 ] corrupted by the same coffee shop babble noise [ 18 ] at SNRs of -5, 0, and 5 dB. T ables 9 & 10 report the results of this process. T esting SNR 5 dB 0 dB -5 dB T raining SNR 5 dB 2.417 1.953 1.442 0 dB 2.378 2.025 1.571 -5 dB 2.283 2.026 1.619 (Noisy: 1.874) (Noisy: 1.471) (Noisy: 1.182) T able 9 . This table presents PESQ test results across dif ferent SNRs for training on one speaker , fine-tuning on a new speaker , and then testing on that new speak er . T esting SNR 5 dB 0 dB -5 dB T raining SNR 5 dB 21.791% 45.588% 87.166% 0 dB 28.342% 40.909% 76.337% -5 dB 58.690% 79.813% 81.684% (Noisy: 33.556%) (Noisy: 72.326%) (Noisy: 94.652%) T able 10 . This table presents WER test results across different SNRs for training on one speaker , fine-tuning on a new speaker , and then testing on that new speak er . When comparing T ables 5 & 6 with T ables 9 & 10, it is noticed that performance on the new speaker now does a good job tracking the performance of a system trained on a speaker then tested on that same speaker . This helps verify the initial hypothesis and it can be concluded that the speech enhancement system trained at 0 dB is able to generalize to new speakers (via fine-tuning) and is markedly rob ust across different SNRs. 5. CONCLUSIONS & FUTURE WORK A fully con volutional neural network based end-to-end speech enhancement system that serv es as a solution to the f amous cocktail party problem has been presented. An ability to gener- alize to new speak ers is presented by fine-tuning of the system with limited data. T est results sho w that the system is rob ust to different babble noise en vironments of varying SNRs. This speech enhancement system shows promising results objec- tiv ely , using PESQ and WER measures, and subjectiv ely by listening to the filtered audio. A few questions to consider for future research pertaining to this speech enhancement system are presented below: 1) What is the optimal model complexity for the task? 2) Does the system continue to generalize well to new en viron- ments and speakers? 3) What is the minimal computational/storage comple xity needed to employ this system in real-time? 6. REFERENCES [1] E. Healy , et al., “ An Algorithm to Improv e Speech Recog- nition in Noise for Hearing-Impaired Listeners. ” The Jour - nal of the Acoustical Society of America , vol. 134, no. 4, 2013, pp. 30293038., doi:10.1121/1.4820893. [2] H. Dillon, Hearing Aids . Thieme, 2012. [3] J. Mcdermott, “The Cocktail Party Prob- lem. ” Curr ent Biology , v ol. 19, no. 22, 2009, doi:10.1016/j.cub .2009.09.005. [4] C. Cherry , “Some Experiments on the Recognition of Speech, with One and with T wo Ears” (PDF). The J our- nal of the Acoustical Society of America ., 1953, 25 (5): 97579. doi:10.1121/1.1907229. [5] Simpson, and Andrew J. R., “Deep T ransform: Cock- tail Party Source Separation via Complex Con vo- lution in a Deep Neural Network. ” 12 Apr . 2015, arxiv .org/abs/1504.02945 [6] K. Paliwal, et al., “The Importance of Phase in Speech Enhancement. ” Speech Communication , vol. 53, no. 4, 2011, pp. 465494., doi:10.1016/j.specom.2010.12.003. [7] T . Sainath, R. W eiss, et al., “Learning the speech front- end with raw wa veform cldnns. ” In Sixteenth Annual Confer ence of the International Speech Communication Association , 2015. [8] G. T rigeorgis, F . Ringev al, et al., “ Adieu features? End- to-end speech emotion recognition using a deep con vo- lutional recurrent network. ” In Acoustics, Speech and Signal Pr ocessing (ICASSP) , 2016 IEEE International Conference, pp. 52005204. IEEE, 2016. [9] J. Thickstun, Z. Harchaoui, and S. Kakade, “Learn- ing features of music from scratch”. arXi v preprint arXiv:1611.09827, 2016. [10] S. Fu, et al, “Raw W aveform-Based Speech Enhance- ment by Fully Con volutional Networks. ” 2017 Asia- P acific Signal and Information Pr ocessing Association Annual Summit and Conference (APSIP A ASC) , 2017, doi:10.1109/apsipa.2017.8281993. [11] Y . Gong and C. Poellabauer, “How do deep con volutional neural networks learn from raw audio wav eforms?”, 2018, A vailable: https://openrevie w .net/forum?id=S1Ow eRb [12] J. McCharthy , A History of the F our Georg es ., V ol. 1, Audiobook A vailable: librivox.or g/a-history-of-the- four-geor ges-volume-1-by-justin-mccarthy/. Narrator: Pamela Nagami [13] “B AR CRO WD for 3 Hours Sound Effect. ” Y ouT ube , 10 Oct. 2013, www .youtube.com/watch?v=bCnJoaXYFGg. [14] “The Noise Cafe 2 hours. ” Y ouT ube , 28 Apr . 2012, www .youtube.com/watch?v=KZV9FmHOsRg. [15] I. Goodfellow , et al., Deep Learning . MIT Press, 2016. A vailable online: http://www .deeplearningbook.org/ [16] D. P . Kingma and J. Ba, “ Adam: A Method for Stochastic Optimization, ” CoRR, vol. abs/1412.6980, 2014. [On- line]. A vailable: http://arxi v .org/abs/1412.6980 [17] S. Ioffe and C. Szegedy , “Batch Normalization: Accel- erating Deep Network T raining by Reducing Internal Cov ariate Shift, ” CoRR, v ol. abs/1502.03167, 2015. [On- line]. A vailable: http://arxi v .org/abs/1502.03167 [18] “Coffee Shop Sounds Background Noise For W ork. ” Y ouT ube , 23 Mar . 2016, www .youtube.com/watch?v=jBNoFk03vWk. [19] Rix, A.W ., et al., “Perceptual ev aluation of speech qual- ity (PESQ) - a new method for speech quality assess- ment of telephone networks and codecs. ” 2001 IEEE International Conference on Acoustics, Speech, and Signal Pr ocessing. Pr oceedings (Cat. No.01CH37221) , doi:10.1109/icassp.2001.941023. [20] A. Morris, et al., “From WER and RIL to MER and WIL: improv ed ev aluation measures for connected speech recognition. ” Institute of Phonetics at Saarland Uni ver- sity , Germany . [21] M. T wain, The $ 30,000 Bequest and Other Stories ., Audiobook A vailable: libri vox.org/30000-bequest-and- other-stories-by-mark-tw ain/, Narrator: T ricia G.