Search Results

In this paper, an approach to analyze voice recognition data to understand how customers use voice recognition systems is explored. The analysis will help identify ASR failures and usability related issues that customers encounter while using the voice recognition system. This paper also examines the impact of these failures on the individual speech domains (media control, phone, navigation, etc.). Such information can be used to improve the current voice recognition system and direct the design of future systems. Infotainment system logs, audio recordings of the voice interactions, their transcriptions and CAN bus data were identified to be rich sources of data to analyze voice recognition usage. Infotainment logs help understand how the system interpreted or responded to customer commands and at what confidence level.

The work presented here is part of the research done in the field of voice biometrics. This paper helps to understand the state-of-the-art in speaker recognition technology potentially capable of solving challenges related to speaker identification (to identify a speaker among multiple speakers) and speaker verification/authentication (to recognize the current speaking person at a pre-defined access level and authenticate accordingly). The research was focused on performing an unbiased evaluation of two individual voice biometric services. The level of accuracy in identifying and authenticating individuals using these services provides an insight into the current state of technology and the state of what other dual authentication methods could be used to achieve a desired True Acceptance Rate (TAR) and False Acceptance Rates (FAR).

Hands-free phone use is the most utilized use case for vehicles equipped with infotainment systems with external microphones that support connection to phones and implement speech recognition. Critically then, achieving hands-free phone call quality in a vehicle is problematic due to the extremely noisy nature of the vehicle environment. Noise generated by wind, mechanical and structural, tire to road, passengers, engine/exhaust, HVAC air pressure and flow are all significant contributors and sources of noise. Other factors influencing the quality of the phone call include microphone placement, cabin acoustics, seat position of the talker, noise reduction of the hands-free system, etc. This paper describes the work done to develop procedures and metrics to quantify the effects that influence the hands-free phone call quality.

Many engine tick and knock issues are clearly audible, yet cannot be characterized by common sound quality metrics such as time-varying loudness, sharpness, fluctuation strength, or roughness. This paper summarizes the recent development and application of an objective metric that agrees with subjective impressions of impulsive engine noise. The metric is based on a general impulsive noise model [1], consisting of a psychoacoustic processing stage followed by a transient detection stage. The psychoacoustic stage is extracted from portions of a time-varying loudness model. The primary output of the impulsive engine noise model is a time series that indicates the location and “intensity” of impulsive engine noise events. The information in this time series is reduced either to a single number metric, or to a frequency-based vector of numbers that indicates the amount of impulsiveness in the recorded sound.

Customer annoyance of steady-state wind noise correlates well with loudness. A common objective metric to capture average loudness is the ISO532B or Zwicker method. However, it has been shown previously that time-varying wind noise can also significantly affect customer annoyance, independent of average loudness. Causes of time-varying wind noise include wind buffeting generated by other vehicles, and also wind gusting. This paper summarizes the development of an objective metric that correlates well with subjective impressions of wind gusting/buffeting. The model is based on a general impulsive noise model with parameters tuned specifically for time-varying wind characteristics. The model consists of a psychoacoustic processing stage followed by a gusting detection stage, where the psychoacoustic stage is extracted from a time-varying loudness model. The output of the gusting model is a time series that indicates the location and “intensity” of wind gusts.

It is well known that customer perception of the annoyance of steady-state wind noise can be fairly well characterized by calculating the loudness of such sounds. Commonly used is the ISO532B or Zwicker method [1]. What is not known, however, is how a customer would react to time-varying wind noise. Such situations can occur when a vehicle experiences cross-wind conditions on the highway. Turbulent air flow generated by either a passing vehicle or when traveling in the wake of another vehicle can cause the wind noise to take on time-varying characteristics. The time-varying wind noise created by such situations is commonly referred to as “buffeting.” Customer complaint field data indicates that wind buffeting is a source of annoyance, but the level of the effect has never been quantified. In this study, binaural sounds were recorded inside an aeroacoustic wind tunnel. Varying degrees of buffeting were simulated using a “blocker” vehicle situated in front of the test vehicle.

In this paper, a systems engineering approach is explored to evaluate the effect of design parameters that contribute to the performance of the embedded Automatic Speech Recognition (ASR) engine in a vehicle. This includes vehicle designs that influence the presence of environmental and HVAC noise, microphone placement strategy, seat position, and cabin material and geometry. Interactions can be analyzed between these factors and dominant influencers identified. Relationships can then be established between ASR engine performance and attribute performance metrics that quantify the link between the two. This helps aid proper target setting and hardware selection to meet the customer satisfaction goals for both teams.

This paper describes two case studies in which multiple microphone processing (beamforming) and microphone location were evaluated to determine their impact on improving embedded automatic speech recognition (ASR) in a vehicle hands-free environment. While each of these case studies was performed using slightly different evaluation set-ups, some specific and general conclusions can be drawn to help guide engineers in selecting the proper microphone location and configuration in a vehicle for the improvement of ASR. There were some outcomes that were common to both dual microphone solutions. When considering both solutions, neither was equally effective across all background noise sources. Both systems appear to be far more effective for noise conditions in which higher frequency energy is present, such as that due to high levels of wind noise and/or HVAC (heating, ventilation and air conditioning) blower noise.

This paper describes a method to validate in-vehicle speech recognition by combining synthetically mixed speech and noise samples with batch speech recognition. Vehicle cabin noises are prerecorded along with the impulse response from the driver's mouth location to the cabin microphone location. These signals are combined with a catalog of speech utterances to generate a noisy speech corpus. Several factors were examined to measure their relative importance on speech recognition robustness. These include road surface and vehicle speed, climate control blower noise, and driver's seat position. A summary of the main effects from these experiments are provided with the most significant factors coming from climate control noise. Additionally, a Signal to Noise Ratio (SNR) experiment was conducted highlighting the inverse relationship with speech recognition performance.