Having participated in a few projects aiming to build simple ASR solutions using CMU Sphinx and open-source models from VoxForge (adapted by myself), I am quite disappointed in their performance. While exposing an acceptable performance in the lab, open source solutions often fail when facing a real-life environment. Even the simplest task like recognition according to a small grammar becomes a problem.
Sure, ASR is a very challenging task that requires huge amounts of dedicated acoustic data, powerful noise cancellation and voice activity detection, as well as careful design of domain-specific language models. But even if those conditions are met, the result is far behind the expectations even for small-vocabulary tasks.
A friend of mine, who was trying to build an IVR system based on Asterisk + UniMRCP + Sphinx, told that the main lesson he has learned was 'not to use the open source technology'. The website of SpokenTech Inc. (http://spokentech.com/) looks abandoned, their demos from speechapi.com do not work.
Can anyone share their experience in development of real-life products upon Sphinx or any other open source engine? In which cases is it worth it to invest your time in that?
I'm sorry to hear that you had difficulties in getting the software to work for you. One of the downsides of open source is that you don't have a support organization there to help out, like you would for a commercial product.
On the other hand there's usually a user community that's willing to help and that over time has accumulated a fair amount of practical advice in its archives. cmusphinx on sourceforge has a very active community and very helpful people, in particular Nickolay Shmyrev.
One thing that often happens is that it's easy to under-estimate the amount of time it takes to get something properly working. People in the field claim that it takes about a person-year of effort to bring up a system in a new domain, and a background in ASR helps. I know that Sphinx is the basis for a number of successful commercial ventures, so it can be made to work.
If you're not altogether discouraged I would urge you to interact with the cmusphinx community and give it another try.
First of all, I would like to thank you for your work: Sphinx is a great piece of software.
Also, without Nickolay's help I would not be able to implement a lot of things.
Indeed, I am quite satisfied with the support of the community. Most questions are already addressed on forums, and my own expertise always helps me in solving the rest.
However, my point is rather about purely technical aspects. For instance, in telephone-based IVR systems the bottleneck is often VAD and noise cancellation. However, Sphinx and UniMRCP do not provide front-end tools to address this issue. Well, Sphinx4 does have VAD (SpeechClassifier + SpeechMarker), but pocketsphinx plugin for UniMRCP does not.
Also, the methods of acoustic and language modelling are pretty standard: MFCCs/PLPs, HMM-GMM, N-grams. While these are good enough for majority of the tasks, we know that e.g. PNCC features are more robust against noise and DNNs can significantly improve the spontaneous speech recognition.
I have a feeling that those questions are better addressed in commercial ASR solutions (like Nuance), while open source community simply does not have enough resources to implement all the cutting edge techniques from the speech recognition area.
Can you as the expert comment on this? Maybe I overestimate the industrial toolkits?
Dear Dmytro,
I'm using CMU Sphinx for a while and I'm very satisfied with its performance and flexibility. I agree that the project is not sufficiently mature to be used out-of-the-box for any application, but the fact that the community (and especially Nicolay) is so willing to help everyone with their application is a huge advantage.
Regarding PNCCs (called "denoise" features in Sphinx): they are implemented and used by default since june 2013.
Regarding VAD: to the best of my knowledge, at the moment pocketsphinx benefits from a new, improved version of VAD, which will be implemented soon in Sphinx4 also.
I found the Denoise class in the Sphinx package, but is it plugged in by default in any configuration? Or I must explicitely include it into the processing pipeline?
I can see it is some kind of adaptation of the PNCC idea to the existing pipeline. Do you know whether the usage of Denoise + MFCC leads to the same effect as the usage of PNCC features? I can find almost no information about Denoise in internets...
- Badges
- Science topic