Input Analysis

Operating on smartphone during driving is dangerous. To let the drivers concen- trate on their driving when using the DriverTalk system, we adopt voice input in DriverTalk. Drivers are allowed to talk in a natural way, i.e. speech, without any hand operation on the smartphone.

To discriminate the talking over DriverTalk from regular talking that happens inside the vehicle, e.g. a driver is talking with a passenger, we define a foreword, “OK Driver”, for the system, just like the way Google Glass is working. When a driver wants to talk to other drivers over DriverTalk, she says “OK Driver” first, and then expresses what she wants to convey to surrounding drivers. The system must recognize the foreword, and thereby understand what the driver is talking and whom the driver is talking to. Fig. 3.4 shows the process of analyzing driver’s input in DriverTalk.

Figure 3.4: Process of input analysis in DriverTalk. This phase extracts the keywords and infers the target of talking.

phone platforms, for instance Siri on iOS. In DriverTalk, we use Android SpeechRec- ognizer to recognize what the driver says.

DriverTalk uses wireless communication to send what a driver says to another driver. If the system directly transmits driver’s speech, the data size will be too large (could be tens to hundreds KBs even for a several-second sentence) to achieve reliable and efficient performance. DriverTalk converts driver’s speech into text through SpeechRecognizer. Later, it transmits text instead of acoustic content to the receivers. In this way, the communication overhead is significantly reduced.

To ease the driver’s effort when talking to the drivers around her and make the speech recognition module work more efficiently, we define a set of keywords,

which map to a set of pre-defined common messages2_{. The keywords and their}

mapping to the pre-defined messages, the corresponding target vehicles of talking when using the messages are listed in Table 3.1. With this definition, instead of saying a long sentences (3rd column in the table), drivers only need to speak the keywords, which could improve the performance of speech recognition.

After recognizing the keywords defined in the table (2nd column), if the key- words are associated with some pre-defined message, DriverTalk fetches the in- dex of the corresponding pre-defined message. The system only needs to transmit the index instead of the complete message to the receiver. At the receiver side,

2_{In this work, based on the context in narration, the word “message” could be one of the two}

kinds of meaning: a) The content of talking. b) The communication/network term for data packet, which includes network address, talking content and other necessary information.

Table 3.1: Mapping of Keywords–Predefined Message–Target Vehicle

Index Keyword Message Target Vehicle(s) of Talking

1 Change

Left

I want to change to the left lane.

Vehicles on left lane moving par-

allel and behind (at position _¯

±).

2 Change

Right

I want to change to the right lane.

Vehicles on right lane moving parallel and behind (at position

° ³).

3 Tailgate Don’t tailgate me. Vehicle moving behind on the

same lane (at position_²).

4 Too

Slow

You are moving too slow and blocking the traffic.

Vehicle moving in front of ego- vehicle on the same lane (at po-

sition_­).

5 Lane

Closed

This lane is closed

ahead.

Vehicle moving behind on the

same lane (at position_²).

6 Accident There is accident

ahead.

All vehicles moving behind on the same or adjacent lanes (at

position_{± ² ³}).

7 Problem There is something

wrong with my car.

All vehicles moving around.

8 Slow

Down

I will slow down. Vehicle moving behind on the

same lane (at position_²).

9 Move Please move. Vehicle moving in front of ego-

vehicle on the same lane (at po-

sition_­).

10 Light On Your light is on. Vehicle moving in front of ego-

vehicle on the same lane (at po-

sition_­).

11 Question Question: All vehicles moving around.

12 Answer Answer: The sender of the corresponding

DriverTalk brings back the corresponding message and plays back (which will be described in Section 3.3.D). At the same time, the system infers which target vehi- cle the ego-driver is talking to based on the relationship depicted in this table. The knowledge of target vehicle’s position is used to select proper VID for the message to be transmitted, which will be clarified in Section 3.3.B.

Different from the regulated messages defined as 1∼10 in Table 3.1, the key-

words “Question” and “Answer” give more flexibility, which could be used to ask questions and answer questions. Once DriverTalk detects one of these two key- words, it converts all what the driver says after the keyword into text as the content of question or answer. The target of a question message is all vehicles moving around the questioner, while the answer message is only targeted at the corre- sponding questioner.

In the future, we will define more pre-defined messages and possibly provide drivers the opportunity to talk in a more flexible way.