Call2 Monitoring BABIR

Public Summary Month 8/2012

Wed, 26 Sep 2012 08:21:30 GMT

Vocally and Aldebaran has provided a first version of the functions of echo cancellation, noise reduction and sound localization. They have built a corpus of audio signals that will be used to model the audio system of Nai and to evaluate the performances of the developed functions. The integration on Nao was delayed because it is not possible to make a synchronized acquisition of the input and the output audio buffers. This synchronization is necessary for the echo cancellation function in which the robot is supposed to cancel its own voice in the audio signal acquired by its microphones. Aldebaran has developed standalone software to overcome this problem. It will allow the testing of the Vocally function.

Public Summary Month 5/2012

Fri, 08 Jun 2012 14:52:22 GMT

Vocally and Aldebaran still work on improving their own echo cancellation, noise reduction and sound localization algorithms. Since few time, a new speech recognition engine has been integrated on Nao, which allows to have 95% of recognition rate when the speaker is saying words at 1m of the robot, even with important fan noise. It has been decided that the recognition of the short orders said to the robot will be made directly with this embedded recognition engine. To improve the natural language understanding by the robot, a recognition engine based on a remote server will be used. The improvments achieved by vocally on the performances of this engine will soon be evaluated.

Public Summary Month 3/2012

Wed, 28 Mar 2012 13:43:36 GMT

Vocally has been working on echo cancellation (the robot listens while it is talking). The implementation on Nao raised two problems. When Nao talks, it saturates the loud speaker and the acquisition of the signals from the loudspeaker and from the microphones is not synchronized. Aldebaran has to work on these issues. Aldebaran has developed an experimental protocol to have an objective way of evaluation the performances of the development made within the project.

Public Summary Month 1/2012

Tue, 14 Feb 2012 16:12:50 GMT

Aldebaran has integrated a first version of the noise filter initially available at Vocally. The performances obtained by this version will be a reference to evaluate the improvement brought by the project. Using a new version of Nao supplied by Aldebaran, Vocally has built an audio corpus that will be used to optimize the low level processing: the noise filter, the echo cancellation function and the robust recognition of the word Nao.

Public Summary Month 11/2011

Fri, 09 Dec 2011 16:49:09 GMT

Aldebaran has tested the performances of the audio functions on the new head of Nao integrating an Intel ATOM CPU. The noise filter delivered by Vocally to Aldebaran uses 4% of CPU. Aldebaran evaluated the performances of the localization of the speaker with the ATOM. It uses 5% of CPU localize the speaker at a precision of 20°. Vocally will now optimize his software on the new head and Aldebaran will evaluate the improvement of the localization brought by the noise filter.

Public Summary Month 09/2011

Mon, 31 Oct 2011 11:05:53 GMT

Vocally and Aldebaran have worked on the integration of Vocally software elements, which aims to improve the robot Nao’s audition by noise reduction and voice print recognition, into the current hardware architecture Nao. Efforts have been mainly dedicated to the assessment of the noise reduction engine developed by Vocally. To measure the effects that Vocally noise reduction has on speech recognition engines currently embedded into the robot, it seems already very useful to remove ambient background noise, starting with the more dominant noise source into NAO's head : its fan.

Public Summary Month 7/2011

Wed, 27 Jul 2011 17:02:30 GMT

In the BABIR experiment, Aldebaran Robotics and Vocally aim at improving the quality of interaction between a human and a domestic robot thanks to the introduction of low level audio processing. Vocal interface (speech and listening) is the most intuitive way to communicate between people. That is the reason why the co-worker or companion robot needs a good audition and a good voice. Of course, the primary sense for a robot is vision that allows action. Many developments have been done in this domain and a lot has still to be done. But if we consider that the new robotic generation has to be closer and closer to the human, the communication become more and more important. So it is necessary that robots can speak and, above all, listen.

Today, speech recognition units are not satisfactory as soon as the acoustic environment is not perfectly controlled. The challenge we take up is to develop a speech recognition unit that works in a very robust way on a 57cm moving humanoid robot, Nao, with a limited embedded computing power and who can be up to 4m from the speaker in a domestic and possibly noisy environment.

BABIR partners work currently on the specifications. The scenario on which they agreed is the following : the Nao robot is walking around when someone calls it. The robot is able to distinguish the call in spite of the surrounding noise and its internal noises (fan, mechanical noises, and collisions of the feet on the ground when walking).

A list of words or expressions to be recognized by Nao both in French and in English was fixed. At the first stage, there are 18 of them, such as Yes, No, Nao, Don’t move, Silence, Higher, Thank you… After recognizing these expressions, Nao will act according to them.