I thought this recent research paper from ARM might be interesting to contestants, especially those looking at running on the Pi:\nhttps://arxiv.org/abs/1711.07128
\n\nIt has accuracy, size, and speed results for a lot of different approaches to running on-device speech models, using the same dataset as this contest.
", "messages": [ { "id": 253801, "postDate": "2017-12-05T17:07:31.543Z", "content": "I thought this recent research paper from ARM might be interesting to contestants, especially those looking at running on the Pi:\nhttps://arxiv.org/abs/1711.07128
\n\nIt has accuracy, size, and speed results for a lot of different approaches to running on-device speech models, using the same dataset as this contest.
", "rawMarkdown": "I thought this recent research paper from ARM might be interesting to contestants, especially those looking at running on the Pi:\nhttps://arxiv.org/abs/1711.07128\n\nIt has accuracy, size, and speed results for a lot of different approaches to running on-device speech models, using the same dataset as this contest.", "votes": 21 }, { "id": 254424, "postDate": "2017-12-06T23:50:59.300Z", "content": "@Pete, thanks for posting our work here.
\n\n@Jeff, all models in Section 4.1 use 40 MFCC features which is the same as used in DNN[5] and CNN[6] papers from Google Research. As discussed in Section 4.3, the number of MFCC features is also used as a hyperparameter for the models. Based on our experiments, we find 10 MFCC features are enough, which is what we used for the architecture exploration and described in the Appendix.
\n\n@Human Analog, we will release all the Tensorflow models and scripts we have used for this paper on github so you can try them yourself š
\n\nWe have put the scripts and models here: https://github.com/ARM-software/ML-KWS-for-MCU
", "rawMarkdown": "@Pete, thanks for posting our work here.\n\n@Jeff, all models in Section 4.1 use 40 MFCC features which is the same as used in DNN[5] and CNN[6] papers from Google Research. As discussed in Section 4.3, the number of MFCC features is also used as a hyperparameter for the models. Based on our experiments, we find 10 MFCC features are enough, which is what we used for the architecture exploration and described in the Appendix.\n\n@Human Analog, we will release all the Tensorflow models and scripts we have used for this paper on github so you can try them yourself š\n\nWe have put the scripts and models here: https://github.com/ARM-software/ML-KWS-for-MCU", "votes": 8, "replies": [ { "id": 254644, "postDate": "2017-12-07T10:07:56.337Z", "content": "Awesome. :-) I ask because several of us have found that a high validation score doesn't necessarily give a high leaderboard score as well. Which would mean either the validation set isn't a good benchmark for real-world performance, of the Kaggle LB test set isn't (or we're doing something else wrong, of course).
", "rawMarkdown": "Awesome. :-) I ask because several of us have found that a high validation score doesn't necessarily give a high leaderboard score as well. Which would mean either the validation set isn't a good benchmark for real-world performance, of the Kaggle LB test set isn't (or we're doing something else wrong, of course)." }, { "id": 256033, "postDate": "2017-12-11T00:58:55.577Z", "content": "Thanks for the info, Liangzhen. I reproduced your experiment on the largest DSCNN model, but it resulted in 94.7% on the test set instead of 95.4%. On the Kaggle leaderboard, it resulted in a score of 0.55. If anyone else here tried reproducing the results from the paper, what were the results you observed?
", "rawMarkdown": "Thanks for the info, Liangzhen. I reproduced your experiment on the largest DSCNN model, but it resulted in 94.7% on the test set instead of 95.4%. On the Kaggle leaderboard, it resulted in a score of 0.55. If anyone else here tried reproducing the results from the paper, what were the results you observed?", "votes": 1 }, { "id": 256372, "postDate": "2017-12-11T20:10:05.357Z", "content": "https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/44250\nHave you read the following discussion about necessary change needed for submitting on Kaggle? LB~0.55 seems to be unreasonable.
", "rawMarkdown": "https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/44250\nHave you read the following discussion about necessary change needed for submitting on Kaggle? LB~0.55 seems to be unreasonable. " }, { "id": 256374, "postDate": "2017-12-11T20:15:39.260Z", "content": "Hi, Ben. Yes, I've read that discussion. With a different model architecture, I was able to get .84 on the leaderboard.
", "rawMarkdown": "Hi, Ben. Yes, I've read that discussion. With a different model architecture, I was able to get .84 on the leaderboard." }, { "id": 256377, "postDate": "2017-12-11T20:42:13.963Z", "content": "that's interesting....didn't expect the discrepancy can be so large. Can you try the smallest DSCNN model to see how it performs? I will train some models to give a shot as well
", "rawMarkdown": "that's interesting....didn't expect the discrepancy can be so large. Can you try the smallest DSCNN model to see how it performs? I will train some models to give a shot as well" }, { "id": 256435, "postDate": "2017-12-11T23:27:43.787Z", "content": "94.7% vs. 95.4% seems reasonable to me considering the different initializations and other factors (Adam vs. SGD, learning rate, step etc.). Also, the largest DS-CNN model (with ~400k weights) resulted in LB score of 0.83. We will be open-sourcing the scripts and models shortly on github.
", "rawMarkdown": "94.7% vs. 95.4% seems reasonable to me considering the different initializations and other factors (Adam vs. SGD, learning rate, step etc.). Also, the largest DS-CNN model (with ~400k weights) resulted in LB score of 0.83. We will be open-sourcing the scripts and models shortly on github.\n", "votes": 5 }, { "id": 256439, "postDate": "2017-12-11T23:38:26.957Z", "content": "Thanks, Liangzhen. What were the training parameter values that you used to reach .83? Iām guessing you changed window size, stride, MFCC bins, etc.
", "rawMarkdown": "Thanks, Liangzhen. What were the training parameter values that you used to reach .83? Iām guessing you changed window size, stride, MFCC bins, etc." }, { "id": 257233, "postDate": "2017-12-13T18:50:30.033Z", "content": "We have put the scripts and models here: https://github.com/ARM-software/ML-KWS-for-MCU
\n\nWe have also included the exact training commands (with all hyperparameters) to generate the models.
", "rawMarkdown": "We have put the scripts and models here: https://github.com/ARM-software/ML-KWS-for-MCU\n\nWe have also included the exact training commands (with all hyperparameters) to generate the models.", "votes": 4 }, { "id": 257386, "postDate": "2017-12-14T06:15:30.843Z", "content": "Thank you for making your code open sourced. I got 0.74 in public score using the pre-trained DNN_S model. This pre-trained model is worse than tutorial's default conv model that gave me 0.78.
", "rawMarkdown": "Thank you for making your code open sourced. I got 0.74 in public score using the pre-trained DNN_S model. This pre-trained model is worse than tutorial's default conv model that gave me 0.78. " }, { "id": 257677, "postDate": "2017-12-14T19:10:56.700Z", "content": "The DNN_S model gives 84.6% on the validation set compared to ~89% of the default conv model. So your observation in LB makes sense here, and DNN_S is a much (~10X) smaller model.
", "rawMarkdown": "The DNN_S model gives 84.6% on the validation set compared to ~89% of the default conv model. So your observation in LB makes sense here, and DNN_S is a much (~10X) smaller model." }, { "id": 257947, "postDate": "2017-12-15T06:00:49.353Z", "content": "I am sorry. I confirmed that your DS_CNN_L model gives 0.83 in LB.
", "rawMarkdown": "I am sorry. I confirmed that your DS_CNN_L model gives 0.83 in LB.", "votes": 2 }, { "id": 260525, "postDate": "2017-12-20T11:52:34.080Z", "content": "@Liangzhen Lai Hi!, is this project only published the model? then what about the data pre-processing before MFCC feature? like sample rate and bit accuracy? and all the workflow before generate MFCC feature? seems that all the micro-controllers have only 12bit accuracy ADC for MIC. will this work?
", "rawMarkdown": "@Liangzhen Lai Hi!, is this project only published the model? then what about the data pre-processing before MFCC feature? like sample rate and bit accuracy? and all the workflow before generate MFCC feature? seems that all the micro-controllers have only 12bit accuracy ADC for MIC. will this work?" }, { "id": 260779, "postDate": "2017-12-20T22:07:53.237Z", "content": "We are planning to release the code as well, check this: https://github.com/ARM-software/ML-KWS-for-MCU/issues/2
\n\nYou are right that most microcontrollers have 12-bit ADC. Some embedded development boards have dedicated audio codec chip to get higher resolution. For example, the demo we show in the paper uses STM32F746G-DISCO and is able to get 16-bit/16,000 sample rate, which is the same as this speech command dataset. We don't have any results for 12-bit, but I guess it shouldn't be too hard to just quantize the 16-bit .wav files into 12-bit and test it out.
", "rawMarkdown": "We are planning to release the code as well, check this: https://github.com/ARM-software/ML-KWS-for-MCU/issues/2\n\nYou are right that most microcontrollers have 12-bit ADC. Some embedded development boards have dedicated audio codec chip to get higher resolution. For example, the demo we show in the paper uses STM32F746G-DISCO and is able to get 16-bit/16,000 sample rate, which is the same as this speech command dataset. We don't have any results for 12-bit, but I guess it shouldn't be too hard to just quantize the 16-bit .wav files into 12-bit and test it out.", "votes": 1 }, { "id": 260849, "postDate": "2017-12-21T03:30:31.767Z", "content": "so when will you release the code of the MFCC feature extracting for STM32-DISCO ?
", "rawMarkdown": "so when will you release the code of the MFCC feature extracting for STM32-DISCO ? " }, { "id": 261185, "postDate": "2017-12-22T00:23:36.287Z", "content": "We are targeting late Jan or early Feb with an end-to-end example. Meanwhile, you can also try to follow tensorflow mfcc implementation (https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/mfcc.cc) with some functions (e.g. fft and cosine) from CMSIS-DSP.
", "rawMarkdown": "We are targeting late Jan or early Feb with an end-to-end example. Meanwhile, you can also try to follow tensorflow mfcc implementation (https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/mfcc.cc) with some functions (e.g. fft and cosine) from CMSIS-DSP." }, { "id": 262350, "postDate": "2017-12-26T09:49:27.027Z", "content": "It seems imposiable to do fft on a cheap MCU(below 100MHz) in realtime, so this guy made a uSpeech (https://github.com/arjo129/uSpeech) to take place fft&mfcc. but it cant use the model you provided. waitting for you end-to-end solution...
", "rawMarkdown": "It seems imposiable to do fft on a cheap MCU(below 100MHz) in realtime, so this guy made a uSpeech (https://github.com/arjo129/uSpeech) to take place fft&mfcc. but it cant use the model you provided. waitting for you end-to-end solution..." }, { "id": 264361, "postDate": "2018-01-02T23:01:00.177Z", "content": "@Liangzhen, how to apply pre-trained model to LB test set?
", "rawMarkdown": "@Liangzhen, how to apply pre-trained model to LB test set?" }, { "id": 264804, "postDate": "2018-01-03T22:51:22.620Z", "content": "You can refer to the end of README.md. Use label_wav.py to generate the prediction for each file in the LB test set.
", "rawMarkdown": "You can refer to the end of README.md. Use label_wav.py to generate the prediction for each file in the LB test set." } ] }, { "id": 254347, "postDate": "2017-12-06T19:16:59.110Z", "content": "Nice paper! It's pretty amazing they're running on MCUs down to Cortex M0s. The MobileNet implementation with separable convolutions might be worth a look, I found a keras implementation on github. I used to work on semiconductor design and we used M0s and M4s. M4s have a Harvard architecture designed for FFTs IIRC, M0s only have a single AHB interface for both instruction and data. I never imagined they'd be used for dense matrix multiplies!
\n\nBill Dally (NVIDA / Stanford) has some cool papers about stochastic rounding of weights, network pruning, and storing weights in a LUT to avoid moving so much data around. Those would all help in fettling a larger model for an embedded target. Here's a great 2-hour talk from NIPS 2015: https://youtu.be/J-GOkwiwg4c
\n\nI just need another couple of months to explore these techniques for the Raspberry Pi :-) !
", "rawMarkdown": "Nice paper! It's pretty amazing they're running on MCUs down to Cortex M0s. The MobileNet implementation with separable convolutions might be worth a look, I found a keras implementation on github. I used to work on semiconductor design and we used M0s and M4s. M4s have a Harvard architecture designed for FFTs IIRC, M0s only have a single AHB interface for both instruction and data. I never imagined they'd be used for dense matrix multiplies!\n\nBill Dally (NVIDA / Stanford) has some cool papers about stochastic rounding of weights, network pruning, and storing weights in a LUT to avoid moving so much data around. Those would all help in fettling a larger model for an embedded target. Here's a great 2-hour talk from NIPS 2015: https://youtu.be/J-GOkwiwg4c \n\nI just need another couple of months to explore these techniques for the Raspberry Pi :-) !", "votes": 1 }, { "id": 254156, "postDate": "2017-12-06T10:21:06.353Z", "content": "Interesting! But how well do these models score on the Kaggle leaderboard? ;-)
", "rawMarkdown": "Interesting! But how well do these models score on the Kaggle leaderboard? ;-)", "votes": 2 }, { "id": 301438, "postDate": "2018-03-22T18:31:43.127Z", "content": "Working on decision trees / random forests for use on microcontrollers here: https://github.com/jonnor/emtrees\nMaybe I'll test it on keyword spotting in the future.
", "rawMarkdown": "Working on decision trees / random forests for use on microcontrollers here: https://github.com/jonnor/emtrees\nMaybe I'll test it on keyword spotting in the future." }, { "id": 260519, "postDate": "2017-12-20T11:41:20.427Z", "content": "@Liangzhen Lai Hi!, is this project only published the model? then what about the data pre-processing before MFCC feature? like sample rate and bit accuracy? and all the workflow before generate MFCC feature? seems that all the micro-controllers have only 12bit accuracy ADC for MIC. will this work?
", "rawMarkdown": "@Liangzhen Lai Hi!, is this project only published the model? then what about the data pre-processing before MFCC feature? like sample rate and bit accuracy? and all the workflow before generate MFCC feature? seems that all the micro-controllers have only 12bit accuracy ADC for MIC. will this work?" }, { "id": 254345, "postDate": "2017-12-06T19:02:02.513Z", "content": "Thanks for the link to the paper, Pete. I noticed that the paragraph under appendix A states that 10 MFCC features were used, but section 4.1 \"Training Results\" states that 40 MFCC features were used. Is that a typo in appendix A? Is 40 MFCC features typical in speech recognition systems?
", "rawMarkdown": "Thanks for the link to the paper, Pete. I noticed that the paragraph under appendix A states that 10 MFCC features were used, but section 4.1 \"Training Results\" states that 40 MFCC features were used. Is that a typo in appendix A? Is 40 MFCC features typical in speech recognition systems?", "replies": [ { "id": 254617, "postDate": "2017-12-07T09:00:39.310Z", "content": "yes, 40 feature dim is more common in SR system.
", "rawMarkdown": "yes, 40 feature dim is more common in SR system." } ] }, { "id": 459172, "postDate": "2019-01-21T10:34:53.970Z", "rawMarkdown": "", "isDeleted": true } ], "comments": [ { "id": 254424, "author_name": "Liangzhen Lai", "author_url": "", "post_date": "2017-12-06T23:50:59.300000", "content": "@Pete, thanks for posting our work here.
\n\n@Jeff, all models in Section 4.1 use 40 MFCC features which is the same as used in DNN[5] and CNN[6] papers from Google Research. As discussed in Section 4.3, the number of MFCC features is also used as a hyperparameter for the models. Based on our experiments, we find 10 MFCC features are enough, which is what we used for the architecture exploration and described in the Appendix.
\n\n@Human Analog, we will release all the Tensorflow models and scripts we have used for this paper on github so you can try them yourself š
\n\nWe have put the scripts and models here: https://github.com/ARM-software/ML-KWS-for-MCU
", "votes": 8, "replies": [ { "id": 254644, "author_name": "Human Analog", "author_url": "", "post_date": "2017-12-07T10:07:56.337000", "content": "Awesome. :-) I ask because several of us have found that a high validation score doesn't necessarily give a high leaderboard score as well. Which would mean either the validation set isn't a good benchmark for real-world performance, of the Kaggle LB test set isn't (or we're doing something else wrong, of course).
", "votes": 0, "replies": [] }, { "id": 256033, "author_name": "Jeff", "author_url": "", "post_date": "2017-12-11T00:58:55.577000", "content": "Thanks for the info, Liangzhen. I reproduced your experiment on the largest DSCNN model, but it resulted in 94.7% on the test set instead of 95.4%. On the Kaggle leaderboard, it resulted in a score of 0.55. If anyone else here tried reproducing the results from the paper, what were the results you observed?
", "votes": 1, "replies": [] }, { "id": 256372, "author_name": "BenZhang", "author_url": "", "post_date": "2017-12-11T20:10:05.357000", "content": "https://www.kaggle.com/c/tensorflow-speech-recognition-challenge/discussion/44250\nHave you read the following discussion about necessary change needed for submitting on Kaggle? LB~0.55 seems to be unreasonable.
", "votes": 0, "replies": [] }, { "id": 256374, "author_name": "Jeff", "author_url": "", "post_date": "2017-12-11T20:15:39.260000", "content": "Hi, Ben. Yes, I've read that discussion. With a different model architecture, I was able to get .84 on the leaderboard.
", "votes": 0, "replies": [] }, { "id": 256377, "author_name": "BenZhang", "author_url": "", "post_date": "2017-12-11T20:42:13.963000", "content": "that's interesting....didn't expect the discrepancy can be so large. Can you try the smallest DSCNN model to see how it performs? I will train some models to give a shot as well
", "votes": 0, "replies": [] }, { "id": 256435, "author_name": "Liangzhen Lai", "author_url": "", "post_date": "2017-12-11T23:27:43.787000", "content": "94.7% vs. 95.4% seems reasonable to me considering the different initializations and other factors (Adam vs. SGD, learning rate, step etc.). Also, the largest DS-CNN model (with ~400k weights) resulted in LB score of 0.83. We will be open-sourcing the scripts and models shortly on github.
", "votes": 5, "replies": [] }, { "id": 256439, "author_name": "Jeff", "author_url": "", "post_date": "2017-12-11T23:38:26.957000", "content": "Thanks, Liangzhen. What were the training parameter values that you used to reach .83? Iām guessing you changed window size, stride, MFCC bins, etc.
", "votes": 0, "replies": [] }, { "id": 257233, "author_name": "Liangzhen Lai", "author_url": "", "post_date": "2017-12-13T18:50:30.033000", "content": "We have put the scripts and models here: https://github.com/ARM-software/ML-KWS-for-MCU
\n\nWe have also included the exact training commands (with all hyperparameters) to generate the models.
", "votes": 4, "replies": [] }, { "id": 257386, "author_name": "Tony Y.", "author_url": "", "post_date": "2017-12-14T06:15:30.843000", "content": "Thank you for making your code open sourced. I got 0.74 in public score using the pre-trained DNN_S model. This pre-trained model is worse than tutorial's default conv model that gave me 0.78.
", "votes": 0, "replies": [] }, { "id": 257677, "author_name": "Liangzhen Lai", "author_url": "", "post_date": "2017-12-14T19:10:56.700000", "content": "The DNN_S model gives 84.6% on the validation set compared to ~89% of the default conv model. So your observation in LB makes sense here, and DNN_S is a much (~10X) smaller model.
", "votes": 0, "replies": [] }, { "id": 257947, "author_name": "Tony Y.", "author_url": "", "post_date": "2017-12-15T06:00:49.353000", "content": "I am sorry. I confirmed that your DS_CNN_L model gives 0.83 in LB.
", "votes": 2, "replies": [] }, { "id": 260525, "author_name": "zhouhua", "author_url": "", "post_date": "2017-12-20T11:52:34.080000", "content": "@Liangzhen Lai Hi!, is this project only published the model? then what about the data pre-processing before MFCC feature? like sample rate and bit accuracy? and all the workflow before generate MFCC feature? seems that all the micro-controllers have only 12bit accuracy ADC for MIC. will this work?
", "votes": 0, "replies": [] }, { "id": 260779, "author_name": "Liangzhen Lai", "author_url": "", "post_date": "2017-12-20T22:07:53.237000", "content": "We are planning to release the code as well, check this: https://github.com/ARM-software/ML-KWS-for-MCU/issues/2
\n\nYou are right that most microcontrollers have 12-bit ADC. Some embedded development boards have dedicated audio codec chip to get higher resolution. For example, the demo we show in the paper uses STM32F746G-DISCO and is able to get 16-bit/16,000 sample rate, which is the same as this speech command dataset. We don't have any results for 12-bit, but I guess it shouldn't be too hard to just quantize the 16-bit .wav files into 12-bit and test it out.
", "votes": 1, "replies": [] }, { "id": 260849, "author_name": "zhouhua", "author_url": "", "post_date": "2017-12-21T03:30:31.767000", "content": "so when will you release the code of the MFCC feature extracting for STM32-DISCO ?
", "votes": 0, "replies": [] }, { "id": 261185, "author_name": "Liangzhen Lai", "author_url": "", "post_date": "2017-12-22T00:23:36.287000", "content": "We are targeting late Jan or early Feb with an end-to-end example. Meanwhile, you can also try to follow tensorflow mfcc implementation (https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/mfcc.cc) with some functions (e.g. fft and cosine) from CMSIS-DSP.
", "votes": 0, "replies": [] }, { "id": 262350, "author_name": "zhouhua", "author_url": "", "post_date": "2017-12-26T09:49:27.027000", "content": "It seems imposiable to do fft on a cheap MCU(below 100MHz) in realtime, so this guy made a uSpeech (https://github.com/arjo129/uSpeech) to take place fft&mfcc. but it cant use the model you provided. waitting for you end-to-end solution...
", "votes": 0, "replies": [] }, { "id": 264361, "author_name": "Adam Wang", "author_url": "", "post_date": "2018-01-02T23:01:00.177000", "content": "@Liangzhen, how to apply pre-trained model to LB test set?
", "votes": 0, "replies": [] }, { "id": 264804, "author_name": "Liangzhen Lai", "author_url": "", "post_date": "2018-01-03T22:51:22.620000", "content": "You can refer to the end of README.md. Use label_wav.py to generate the prediction for each file in the LB test set.
", "votes": 0, "replies": [] } ] }, { "id": 254347, "author_name": "tb303", "author_url": "", "post_date": "2017-12-06T19:16:59.110000", "content": "Nice paper! It's pretty amazing they're running on MCUs down to Cortex M0s. The MobileNet implementation with separable convolutions might be worth a look, I found a keras implementation on github. I used to work on semiconductor design and we used M0s and M4s. M4s have a Harvard architecture designed for FFTs IIRC, M0s only have a single AHB interface for both instruction and data. I never imagined they'd be used for dense matrix multiplies!
\n\nBill Dally (NVIDA / Stanford) has some cool papers about stochastic rounding of weights, network pruning, and storing weights in a LUT to avoid moving so much data around. Those would all help in fettling a larger model for an embedded target. Here's a great 2-hour talk from NIPS 2015: https://youtu.be/J-GOkwiwg4c
\n\nI just need another couple of months to explore these techniques for the Raspberry Pi :-) !
", "votes": 1, "replies": [] }, { "id": 254156, "author_name": "Human Analog", "author_url": "", "post_date": "2017-12-06T10:21:06.353000", "content": "Interesting! But how well do these models score on the Kaggle leaderboard? ;-)
", "votes": 2, "replies": [] }, { "id": 301438, "author_name": "Jon Nordby", "author_url": "", "post_date": "2018-03-22T18:31:43.127000", "content": "Working on decision trees / random forests for use on microcontrollers here: https://github.com/jonnor/emtrees\nMaybe I'll test it on keyword spotting in the future.
", "votes": 0, "replies": [] }, { "id": 260519, "author_name": "zhouhua", "author_url": "", "post_date": "2017-12-20T11:41:20.427000", "content": "@Liangzhen Lai Hi!, is this project only published the model? then what about the data pre-processing before MFCC feature? like sample rate and bit accuracy? and all the workflow before generate MFCC feature? seems that all the micro-controllers have only 12bit accuracy ADC for MIC. will this work?
", "votes": 0, "replies": [] }, { "id": 254345, "author_name": "Jeff", "author_url": "", "post_date": "2017-12-06T19:02:02.513000", "content": "Thanks for the link to the paper, Pete. I noticed that the paragraph under appendix A states that 10 MFCC features were used, but section 4.1 \"Training Results\" states that 40 MFCC features were used. Is that a typo in appendix A? Is 40 MFCC features typical in speech recognition systems?
", "votes": 0, "replies": [ { "id": 254617, "author_name": "Feiteng", "author_url": "", "post_date": "2017-12-07T09:00:39.310000", "content": "yes, 40 feature dim is more common in SR system.
", "votes": 0, "replies": [] } ] }, { "id": 459172, "author_name": "", "author_url": "", "post_date": "2019-01-21T10:34:53.970000", "content": "", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "253801": "I thought this recent research paper from ARM might be interesting to contestants, especially those looking at running on the Pi:\nhttps://arxiv.org/abs/1711.07128\n\nIt has accuracy, size, and speed results for a lot of different approaches to running on-device speech models, using the same dataset as this contest.", "254424": "@Pete, thanks for posting our work here.\n\n@Jeff, all models in Section 4.1 use 40 MFCC features which is the same as used in DNN[5] and CNN[6] papers from Google Research. As discussed in Section 4.3, the number of MFCC features is also used as a hyperparameter for the models. Based on our experiments, we find 10 MFCC features are enough, which is what we used for the architecture exploration and described in the Appendix.\n\n@Human Analog, we will release all the Tensorflow models and scripts we have used for this paper on github so you can try them yourself š\n\nWe have put the scripts and models here: https://github.com/ARM-software/ML-KWS-for-MCU", "254347": "Nice paper! It's pretty amazing they're running on MCUs down to Cortex M0s. The MobileNet implementation with separable convolutions might be worth a look, I found a keras implementation on github. I used to work on semiconductor design and we used M0s and M4s. M4s have a Harvard architecture designed for FFTs IIRC, M0s only have a single AHB interface for both instruction and data. I never imagined they'd be used for dense matrix multiplies!\n\nBill Dally (NVIDA / Stanford) has some cool papers about stochastic rounding of weights, network pruning, and storing weights in a LUT to avoid moving so much data around. Those would all help in fettling a larger model for an embedded target. Here's a great 2-hour talk from NIPS 2015: https://youtu.be/J-GOkwiwg4c \n\nI just need another couple of months to explore these techniques for the Raspberry Pi :-) !", "254156": "Interesting! But how well do these models score on the Kaggle leaderboard? ;-)", "301438": "Working on decision trees / random forests for use on microcontrollers here: https://github.com/jonnor/emtrees\nMaybe I'll test it on keyword spotting in the future.", "260519": "@Liangzhen Lai Hi!, is this project only published the model? then what about the data pre-processing before MFCC feature? like sample rate and bit accuracy? and all the workflow before generate MFCC feature? seems that all the micro-controllers have only 12bit accuracy ADC for MIC. will this work?", "254345": "Thanks for the link to the paper, Pete. I noticed that the paragraph under appendix A states that 10 MFCC features were used, but section 4.1 \"Training Results\" states that 40 MFCC features were used. Is that a typo in appendix A? Is 40 MFCC features typical in speech recognition systems?", "459172": "" } }