DARPA Seeks Help With Translation SystemDARPA Seeks Help With Translation System

The Defense Advanced Research Projects Agency (DARPA) is seeking proposals from technology companies to help them develop a high-tech translation system that can accurately detect and translate speech even under noisy conditions.

The Robust Automatic Transcription of Speech program -- which bears the unfortunate acronym "RATS" -- aims to create a system that can accurately determine speech activity, identify a speaker by his or her voice, and pick out keywords even in cases where speech is garbled or where there is noisy interference.

RATS is part of Human Language Technology (HLT) R&D efforts at DARPA, which is the military's experimental research agency.

Firms have until March 29 to enter their proposals into an initial evaluation phase, with July 29 as the drop-dead deadline for submitting proposals, according to a DARPA RFP for the project.

DARPA is soliciting help in three specific technical areas: algorithmic development and signal processing, data collection, and evaluation.

Firms may submit proposals to all three areas, but companies selected for algorithmic development and signal processing cannot be selected for either of the two other technical areas, according to DARPA. Firms also must submit separate proposals for each area.

DARPA anticipates awarding multiple firms contracts for the first technical area, but only one each for the second and third.

The U.S. military has been intercepting and translating enemy data both in times of war and during peacetime to advance intelligence efforts for years, but merely translating speech from another language is no longer serving the best purpose, according to the RFP.

"Existing transcription and translation and speech signal processing technologies are insufficient for working with noisy or degraded speech signals that are of importance to current and future Department of Defense (DoD) operations," according to the RFP.

Specifically, RATS seeks to develop technology that can interpret speech even if the speech signal is distorted. The military currently relies on humans to do this task, but their capabilities are limited when they are listening to multiple channels of speech, according to DARPA.

DARPA is targeting three phases for the project -- a first phase of 18 months, the first six of which will be entirely composed of data gathering; and a second and third phase each of 12 months. At the end of each phase will be a strict evaluation period of the technology, and performance metrics for these periods will get progressively more difficult, according to DARPA.