I am trying to find databases like the LJ Speech Dataset made by Keith Ito. I need to use these datasets in TacoTron 2 (Link), so I think datasets need to be structured in a certain way. the LJ database is linked directly into the tacotron 2 github page, so I think it's safe to assume it's made to work with it. So I think Databases should have the same structure as the LJ. I downloaded the Dataset and I found out that it's structured like this:
main folder:
-wavs
-001.wav
-002.wav
-etc
-metadata.csv: This file is a csv file which contains all the things said in every .wav, in a form like this **001.wav | hello etc.**
So, my question is: Are There other datasets like this one for further training?
But I think there might be problems, for example, the voice from one dataset would be different from the one in one another, would this cause too much problems?
And also different slangs or things like that can cause problems?
There a few resources:
The main ones I would look at are Festvox (aka CMU artic) http://www.festvox.org/dbs/index.html and LibriVoc https://librivox.org/
these guys seem to be maintaining a list
https://github.com/candlewill/Speech-Corpus-Collection
And I am part of a project that is collecting more (shameless self plug): https://github.com/Idlak/Living-Audio-Dataset
Mozilla includes a database of several datasets you can download and use, if you don't need your own custom language or voice: https://voice.mozilla.org/data
Alternatively, you could create your own dataset following the structure you outlined in your OP. The metadata.csv file needs to contain at least two columns -- the first is the path/name of the WAV file (without the .wav extension), and the second column is the text that has been spoken.
Unless you are training Tacotron with speaker embedding/a multi-speaker model, you'd want all the recordings to be from the same speaker. Ideally, the audio quality should be very consistent with a minimum amount of background noise. Some background noise can be removed using RNNoise. There's a script in the Mozilla Discourse group that you can use as a reference. All the recordings files need to be short, 22050 Hz, 16-bit audio clips.
As for slag or local colloquialisms -- not sure; I suspect that as long as the word sounds match what's written (i.e. the phonemes match up), I would expect the system to be able to handle it. Tacotron is able to handle/train on multiple languages.
If you don't have the resources to produce your own recordings, you could use audio from a permissively licensed audiobook in the target language. There's a tutorial on this very topic here: https://medium.com/#klintcho/creating-an-open-speech-recognition-dataset-for-almost-any-language-c532fb2bc0cf
The tutorial has you:
Download the audio from the audiobook.
Remove any parts that aren't useful (e.g. the introduction, foreward, etc) with Audacity.
Use Aeneas to fine-tune and then export a forced alignment between the audio and the text of the e-book, so that the audio can be exported sentence by sentence.
Create the metadata.csv file containing the map from audio to segments. (The format that the post describes seems to include extra columns that aren't really needed for training and are mainly for use by Mozilla's online database).
You can then use this dataset with systems that support LJSpeech, like Mozilla TTS.
Related
I have my twitter archive downloaded and wanted to run word2vec to experiment most similar words, analogies etc on it.
But I am stuck at first step - how to convert a given dataset / csv / document so that it can be input to word2vec? i.e. what is the process to convert data to glove/word2vec format?
Typically implementations of the word2vec & GLoVe algorithms do one or both of:
accept a plain text file, where tokens are delimited by (one or more) spaces, and text is considered each newline-delimited line at a time (with lines that aren't "too long" - usually, short-article or paragraph or sentence per line)
have some language/library-specific interface for feeding texts (lists-of-tokens) to the algorithm as a stream/iterable
The Python Gensim library offers both options for its Word2Vec class.
You should generally try working through one or more tutorials to get a working overview of the steps involved, from raw data to interesting results, before applying such libraries to your own data. And, by examining the formats used by those tutorials – and the extra steps they perform to massage the data into the formats needed by exactly the libraries you're using – you'll also see ideas for how your data needs to be prepared.
I have been trying to pull out financial statements embedded in annual reports in pdf and export them in excel/CSV format using python But I am encountering some problems:
1. A specific Financial statement can be on any page in the report. If I were to process hundreds of pdfs, I would have to specify page numbers which takes alot of time. Is there any way through which the scraper knows where the exact statement is?
2. Some reports span over multiple pages and the end result after scraping a pdf isnt what I want
3. Different annual reports have different financial statement formats. Is there any way to process them and change them to a specific standard format?
I would also appreciate if anyone have done something like this and can share examples.
Ps I am working with python and used tabula and Camelot
I had a similar case where the problem was to extract specific form information from pdfs (name, date of birth and so on). I used the tesseract open source software with pytesseract to perform OCR on the files . Since I did not need the whole pdfs, but specific information from them, I designed an algorithm to find the information: In my case I used simple heuristics (specific fields, specific line number and some other domain specific stuff), but you can also use a machine-learning approach and train a classifier which can find the needed text-parts. You could use domain-specific heuristics as well, because I am sure that a financial statement has special vocabulary or some text markers which indicate its beginning/its end.
I hope I could at least give you some ideas how to approach the problem
P.S.: With tesseract you can also process multipage pdfs. To 3) - Machine learning approach would need some samples to learn a good generalization of how a financial statement may look like.
I have lot of PDF, DOC[X], TIFF and others files (scans from a shared folder). Each file converted into pack of text files: one text file per page.
Each pack of files could contain multiple documents (for example thee contracts). Document kind could be not only contract.
During the processing the pack of the files I don't know what kind of the documents current pack contains and it's possible that one pack contains multiple document kinds (contracts, invoices, etc).
I'm looking for some possible approaches to solve this programmatically.
I'm tried to search something like that but without any success.
UPD: I tried to create binary classificator with scikit-learn and now looking for another solution.
This at its basis, being they are "scans" sounds more like something that could be approached with computer vision, however this is currently far far above my current level of programming.
E.g. projects like SimpleCV may be a good starting point,
http://www.simplecv.org/
Or possibly you could get away with OCR reading the "scans" and working based on the contents. pytesseract seems popular for this type of task,
https://pypi.org/project/pytesseract/
However that still lacks defining how you would tell your program that this part of the image means that this is 3 separate contracts, Is there anything about these files in particular that make this clear, e.g. "1 of 3" on the pages,, a logo or otherwise? that will be the main part that determines how complex a problem you are trying to solve.
Best solution was to create binary classifier (SGDClassifier) and train it on classes first-page and not-first-page. Each item from the dataset was trimmed to 100 tokens (words)
I'm currently working on a little python script to equalize MP3 file.
I've read some docs about MP3 file format (at https://en.wikipedia.org/wiki/ID3)
And i've noticed that in the ID3v2 format there is a field for Equalization (EQUA, EQU2)
Using the python librarie mutagen i've tried to extract theses information from the MP3 but the field isn't present.
What's the right way to equalize MP3 file regardless of the ID3 version ?
Thank in advance. Creekorful
There are two high-level approaches you can take: modify the encoded audio stream, or put metadata on it describing the desired change. Modifying the audio stream is the most compatible, but generally less desirable. However, ID3v1 has no place for this metadata, only ID3v2.2 and up do.
Depending on what you mean by equalize, you might want equalization information stored in the EQA/EQUA/EQU2 frames, or a replay gain volume adjustment stored in the RVA/RVAD/RVA2 frames. Mutagen supports the linked frames, so all but EQA/EQUA. If you need them, it should be straightforward to add them from the information in the actual specification (see 4.12 on http://id3.org/id3v2.4.0-frames). With tests they could likely be contributed back to the project.
Note that Quod Libet, the player paired with Mutagen, has taken a preference for reading and storing replay gain information in a TXXX frame.
I've been developing a tool to compare an audio file recorded on day one to another recorded thirty days later. My training is in linguistics and this tool will be used to catalogue, index, and compare a database of unique vocal recordings. I am aware of commercial grade APIs such as MusicBrainz or EchoNest, but cannot use them for this project. All the files must be locally stored and cannot be contributed to an online database.
At present, I have spectrograms of each file and a batch converter that can convert to almost any sound file. I use a spectrum analyzer to exactly match the spectrograms (like a hash map overlay) and am able to match my results with 96% accuracy. However, as my project grows my storage needs will become far too lofty for this method.
My thought is this - if I can adjust the audio files to a similar frame speed, I should be able to hash code the acoustic data and store the hash strings in a simple SQL table rather than whole audio files or spectrograms. I don't want to hash the whole file - just the acoustics, for matching. I've found a few overkill solutions via Python (dejavu, libmo, etc) but as a linguist, not a computers person, I am unsure if a novice can wrangle the code for hashing audio data
I'm looking to have a way to create hash values (or another checksum) within the next week or so.Thoughts from the interwebz?