• B0_664272
  • 58.7KB
  • zip
  • 0
  • VIP专享
  • 0
  • 2022-04-27 07:19
custom_fe HCopy_UP是bash包装器,用于可与HTK和Kaldi一起使用的自定义前端。 它模仿HTK的HCopy接口,但在内部调用Matlab或Python代码。 这样可以轻松创建以HTK或Kaldi格式编写功能的自定义前端,因此可以将它们合并到可用的多个HTK / Kaldi配方中,而只需进行少量更改即可。 包装器是用bash编写的,因此仅限于unix和cygwin环境。 但是请注意,bash包装器在内部调用了纯Matlab MCopy和纯Python HCopy包装器。 这些也应在Windows中工作。 我创建HCopy_UP作为统一不确定性传播和观察不确定性技术的多个前端的方法。 因此,包装程序即插即用地支持这些技术。 现在可以使用以下示例 STFT-UP用于推导MMSE-MFCC估计量及其剩余不确定性 与CHiME 2013挑战中使用的基于稀疏性的保证 Int
custom_fe ============= HCopy_UP is a bash wrapper for a custom front-end usable with HTK and Kaldi. It imitates HTK's HCopy interface but calls Matlab or Python code internally. This allows to easily create custom front-ends that write the features in HTK or Kaldi format and can be therefore incorporated to the multiple HTK/Kaldi recipes available with minor changes. The wrapper is written in bash, thus limited do unix and cygwin environments. Note however that the bash wrapper calls pure Matlab MCopy and pure Python HCopy wrappers internally. These should also work in Windows. I created HCopy_UP as a way to unify my multiple front-ends for uncertainty propagation and observation uncertainty techniques. The wrappers therefore support these techniques in a plug-and-play fashion. Right now following examples are available * STFT-UP used to derive a MMSE-MFCC estimator and its residual uncertainty * Sparsity based Uncertanties as those used in CHiME 2013 challenge * The DIRHA-GRID front-end for interspeech 2014 You can have a look here for more details Patches are also provided to modify HTK 3.4.1. in order to be able to process these uncertainties using Uncertainty Decoding and Modified Imputation. *Support for Kaldi* The support for Kaldi format features files is already included in the install, via the kaldi-to-matlab toolbox. You just need to set TARGETFORMAT to KALDI in a config file (default being HTK). Note that you need the Kaldi binaries to be available and some other pre-requisites. This is also relatively new so I appreciate feedback for debugging platform issues. There are also tools by other authors that can be combined with this toolbox e.g. integration with Kaldi. Have a look at the Robust Speech Processing Special Interest Group (RoSP) WiKi *Support for Python* Feature extractions are the only part of my code-base that has survived in Matlab despite my switch to Python a while ago. There is already a Python front-end with uncertainty propagation and IMCRA available in this public version, although it has not been tested thoroughly. I expect Python to slowly eat up Matlab's share of code in this repo. **Instalation of Matlab Tools from the zip** If you are familiar with Git, the best way to use these tools is to fork and clone the repo from Github. The other alternative is to download the zip manually (right side of the screen on Github). In that case you can rename the unzipped tools as mv ./custom_fe-master ./custom_fe It should work anyway, it is more a matter of aesthetics. The wrapper makes use of various external toolboxes, Mike Brooke's voicebox toolbox, Emmanuel Vincent's kaldi-to-matlab toolbox and my stft_up_tools and obsunc toolboxes. They are downloaded automatically by using ./custom_fe/install This uses wget, which depending on your platform might not be available. The script will ask you to download them with a browser in this case. Note also that only the needed functions are unzipped. If matlab is available on your bin, this should be enough. If not, you can edit HCopy_UP and set the MATLAB_PATH variable pointing to your binary. **Test Using the GRID-DIRHA Baseline Front-End** As an example, the Matlab front-end for the DIRHA-GRID corpus baseline is here provided. This front-end is also able to read DIRHA-corpora meta-data allowing to perform various oracle knowledge experiments, like e.g. Oracle beamforming or Oracle Voice Activity Detection on DIRHA-corpora, see [1] M. Matassoni, R. F. Astudillo, A. Katsamanis, M. Ravanelli "The DIRHA-GRID corpus: baseline and tools for multi-room distant speech recognition using distributed vmicrophones", Interspeech 2014 Once the Matlab tools are instaled you can do a test run with ./custom_fe/HCopy_UP MAT -C ./custom_fe/MAT/custom/IS2014/config_IS2014 \ ./custom_fe/MAT/stft_up_tools/DATA/s29_pbiz6p.wav \ ./s29_pbiz6p.mfc \ -debug Note that HCopy_UP will indicate the corresponding -debug call using only MCopy.m. This is much faster as it can be done inside Matlab multiple times to debug a custom front-end. Remember to put a breakpoint in MCopy before the exit or it will close Matlab after each run!. **Using Features with Observation Uncertainties** HCopy_UP can be used together with front-ends that produce not only features but also a measure of uncertainty. For example for the front-end above this can be attained using the -up flag together with the config using the MMSE-MFCC estimator (config_IS2014M) as ./custom_fe/HCopy_UP MAT -C ./custom_fe/MAT/custom/IS2014/config_IS2014M \ ./custom_fe/MAT/stft_up_tools/DATA/s29_pbiz6p.wav \ ./s29_pbiz6p.mfc \ -up The additional uncertainties are appended to the normal features resulting in twice the number of features. To process this aditional features with HTK to compute Uncertainty Decoding or Modified Imputation, the patches found [here]( can be used.