As a part of a project, we want to do experiments with synthetic voices where these do not have a singular geographic origin, body, age or gender. We have our own data-set, but I thought of during initial experiments with VCTK and build a voice using Tacotron2 or something similar. Does anyone know if a similar project has been done? Where the physical body that we imagine connected to a voice is intentionally ambiguous. Or other projects where TTS has be trained on a multi-person corpus? Additionally, does anyone know of any caveats or potential problems in terms of this approach? Maybe there could be ways of working with transfer-learning that could be beneficial.
Thanks!
You can check https://github.com/r9y9/deepvoice3_pytorch
Multispeaker samples are available as well as pretrained model you can try.
Related
I have a task: to determine the sound source location.
I had some experience working with tensorflow, creating predictions on some simple features and datasets. I assume that for this task, there would be necessary to analyze the sound frequences and probably other related data on training and then prediction steps. The sound goes from the headset, so human ear is able to detect the direction.
1) Did somebody already perform that? (unfortunately couldn't find any similar project)
2) What kind of caveats could I meet while trying to achieve that?
3) Am I able to do that using this technology approach? Are there any other sound processing frameworks / technologies / open source projects that could help me ?
I am asking that here, since my research on google, github, stackoverflow didn't show me any relevant results on that specific topic, so any help is highly appreciated!
This is typically done with more traditional DSP with multiple sensors. You might want to look into time difference of arrival(TDOA) and direction of arrival(DOA). Algorithms such as GCC-PHAT and MUSIC will be helpful.
Issues that you might encounter are: DOA accuracy is function of the direct to reverberant ratio of the source, i.e. the more reverberant the environment the harder it is to determine the source location.
Also you might want to consider the number of location dimensions you want to resolve. A point in 3D space is much more difficult than a direction relative to the sensors
Using ML as an approach to this is not entirely without merit but you will have to consider what it is you would be learning, i.e. you probably don't want to learn the test rooms reverberant properties but instead the sensors spatial properties.
I'm looking into designing a software platform that will aid linguists and anthropologists in their study of previously unstudied languages. Statistics show that around 1,000 languages exist that have never been studied by a person outside of their respective speaker groups.
My goal is to utilize TensorFlow to make a platform that will allow linguists to study and document these languages more efficiently, and to help them create written systems for the ones that don't have a written system already. One of their current methods of accomplishing such a task is three-fold: 1) Record a native speaker conversing in the language, 2) Listening to that recording and trying to transcribe it into the IPA, 3) From the phonetics, analyzing the phonemics and phonotactics of the language to eventually create a written system for the speaker.
My proposed platform would cut that research time down from a minimum of a year to a maximum of six months. Before I start, I have some questions...
What would be required to train TensorFlow to transcribe live audio into the IPA? Has this already been done? and if so, how would I utilize a previous solution for this project? Is a project like this even possible with TensorFlow? if not, what would you recommend using instead?
My apologies for the magnitude of this question. I don't have much experience in the realm of machine learning, as I am just beginning the research process for this project. Any help is appreciated!
I guess I will take a first shot at answering this. Since the question is pretty general, my answer will have to be pretty general as well.
What would be required. At the very least you would have to have a large dataset of pre-transcribed data. Ideally a large amount of spoken language audio mapped to characters in the phonetic alphabet, so the system could learn the sound of individual characters rather than whole transcribed words. If such a dataset doesn't exist, a less granular dataset could be used, mapping single words to their transcriptions. Then you would need a model, that is the actual neural network architecture implemented in code. And lastly you would need some computing resources. This is not something you can train casually, you would either have to buy some time in a cloud based machine learning framework (like Google Cloud ML) or build a fairly expensive machine to train at home.
Has this been done? I don't know. I don't think so. There have been published papers reporting various degrees of success at training systems to transcribe speech. Here is one, for example, http://deeplearning.stanford.edu/lexfree/lexfree.pdf It seems that since the alphabet you want to transcribe to is specifically designed to capture the way words sound rather than just write down the words you might have more success at training such a model.
Is it possible with TensorFlow. Yes, most likely. TensorFlow is well suited for implementing most modern deep learning architectures. Unless you end up designing some really weird and very original model for this purpose, TensorFlow should work just fine.
Edit: after some thought in part 1, you would have to use a dataset mapping spoken words to their transcriptions, since I expect that the same sound pronounced separately would be different from when the same sound is used in a word.
This has actually been done, albeit in PyTorch, by a group at CMU: https://github.com/xinjli/allosaurus
Recently I stared toying with tensor flow, dnns etc. now I'm trying to implement something more serious, information retrieval from short sentences (doctor instructions).
Unfortunately the dataset I have is, as always, quite "dirty". As I'm trying to use word embeddings, I actually need "clean" data. Take one example:
"Take two pilleach day". There is a missing white space between pill and each. I am implementing "tokenizer improver" to look at each sentence and propose new tokenization based on joint probability of each word in sentence given the frequency of terms in whole document (tf) . As I was doing it today, a thought came to my mind: why bother writing suboptimal solution for this problem when I can employ powerful learning algorithms such as Lstm networks to do that for me. However, as of today, I have only a feeling that it's actually possible to do that. As we know, feelings are not best when it comes to architecting such complex problems. I don't know where to begin: what should be my training set and learning goal.
I know this is a broad question, but I know there are many brilliant people with more knowledge about tensorflow and neural nets, so I'm sure that somebody has either already solved similar problem or just knows how to approach this problem.
Any guidance is welcome, I do not except you to solve this for me of course:)
Besos and all the best to all the tensorflow community:)
Having the same issue. I solved it by using a character level net. Basically I rewrote Character-Aware Neural Language Models, kicked out the whole "words"-elements and just stayed with the caracter level.
Training Data: I took the data I had, as dirty as it was, used the dirty data as targets and made it even more dirty to create inputs.
So your "Take two pilleach day" will be learned as in many cases you do have a clean and similar phrase, e.g. "Take one pill each morning" that with the regime mentioned will serve as target and you train the net on destroyed inputs like "Take oe pileach mornin"
I'm currently working on a project that requires a database categorising websites (e.g. cnn.com = news). We only require broad classifications - we don't need every single URL classified individually. We're talking to the usual vendors of such databases, but most quotes we've had back are quite expensive and often they impose annoying requirements - like having to use their SDKs to query the database.
In the meantime, I've also been exploring the possibility of building such a database myself. I realise that this is not a 5 minute job, so I'm doing plenty of research.
From reading various papers on the subject, it seems a Naive Bayes classifier is generally the standard approach for doing this. However, many of the papers suggest enhancements to improve its accuracy in web classification - typically by making use of other contextual information, such as hyperlinks, header tags, multi-word phrases, the URL, word frequency and so on.
I've been experimenting with Mahout's Naive Bayes classifier against the 20 Newsgroup test dataset, and I can see its applicability to website classification, but I'm concerned about its accuracy for my use case.
Is anyone aware of the feasibility of extending the Bayes classifier in Mahout to take into account additional attributes? Any pointers as to where to start would be much appreciated.
Alternatively, if I'm barking up entirely the wrong tree please let me know!
You can control the input about as much as you'd like. In the end the input is just a feature vector. The feature vector's features can be words, or bigrams -- but they can also be whatever you want. So, yes, you can inject new features by modifying the input as you like.
How best to weave in those features is another topic entirely -- there's not one best way to convert them to numbers. Mahout in Action covers this reasonably well FWIW.
I'm working on some software for children, and looking to add the ability for the software to respond to a number of non-speech sounds. For instance, clapping, barking, whistling, fart noises, etc.
I've used CMU Sphinx and the Windows Speech API in the past, however, as far as I can tell neither of these have any support for non-speech noises, and in fact I believe actively filter them out.
In general I'm looking for "How do I get this functionality" but I suspect it may help if I break it down into three questions that are my guesses for what to search for next:
Is there a way to use one of the main speech recognition engines to recognize non-word sounds by changing an acoustic model or pronunciation lexicon?
(or) Is there already an existing library to do non-word noise recognition?
(or) I have a bit of familiarity with Hidden Markov Models and the underlying tech of voice recognition from college, but no good estimate on how difficult it would be to create a very small noise/sound recognizer from scratch (suppose <20 noises to be recognized). If 1) and 2) fail, any estimation on how long it would take to roll my own?
Thanks
Yes, you can use speech recognition software like CMU Sphinx for recognition of non-speech sounds. For this, you need to create your own acoustical and language models and define the lexicon restricted to your task. But to train the corresponding acoustic model, you must have enough training data with annotated sounds of interest.
In short, the sequence of steps is the following:
First, prepare resources for training: lexicon, dictionary etc. The process is described here: http://cmusphinx.sourceforge.net/wiki/tutorialam. But in your case, you need to redefine phoneme set and the lexicon. Namely, you should model fillers as real words (so, no ++ around) and you don't need to define the full phoneme set. There are many possibilities, but probably the most simple one is to have a single model for all speech phonemes. Thus, your lexicon will look like:
CLAP CLAP
BARK BARK
WHISTLE WHISTLE
FART FART
SPEECH SPEECH
Second, prepare training data with labels: Something similar to VoxForge, but text annotations must contain only labels from your lexicon. Of course, non-speech sounds must be labeled correctly as well. Good question here is where to get large enough amount of such data. But I guess it should be possible.
Having that, you can train your model. The task is simpler compared to speech recognition, for instance, you don't need to use triphones, just monophones.
Assuming equal prior probability of any sound/speech, the simplest language model can be a loop-like grammar (http://cmusphinx.sourceforge.net/wiki/tutoriallm):
#JSGF V1.0;
/**
* JSGF Grammar for Hello World example
*/
grammar foo;
public <foo> = (CLAP | BARK | WHISTLE | FART | SPEECH)+ ;
This is the very basic approach to using ASR toolkit for your task. In can be further improved by fine-tuning HMMs configurations, using statistical language models and using fine-grained phonemes modeling (e.g. distinguishing vowels and consonants instead of having single SPEECH model. It depends on nature of your training data).
Outside the framework of speech recognition, you can build a simple static classifier that will analyze the input data frame by frame. Convolutional neural networks that operate over spectrograms perform quite well for this task.
I don't know any existing libraries you can use, I suspect you may have to roll your own.
Would this paper be of interest? It has some technical detail, they seem to be able to recognise claps and differentiate them from whistles.
http://www.cs.bham.ac.uk/internal/courses/robotics/halloffame/2001/team14/sound.htm