Software Design Study for Voice Interface

Author:Eric Meinhardt

Change Record

2013.08.02 - Draft 0.5 posted.

Introduction

Purpose & Scope

This document is intended to detail for developers of the ERAS voice interface library and other ERAS C3 components how ERAS’s voice interface library is predicted to achieve the features laid out in the system requirements document [1], now and projected into the future as more details about the compatibility of different open source spoken dialogue systems with the broader hardware and software requirements of the ERAS C3 prototype.

Glossary

ERAS
European Mars Analogue Station
ERAS C3
European Mars Analogue Station Command, Control, and Communication
CMU
Carnegie-Mellon University.
ASR
Automatic speech recognition.
SLU
Spoken language understanding.
SDS
Spoken dialogue system.

Design Considerations

Interface requirements and planning with future feature growth in mind are the two chief sources of design considerations:

  • spoken-language command interfaces, as opposed to e.g. systems for coarse- grained understanding of large /text/s or telephone dialog systems for database query or customer service transactions, pose unique and less- studied challenges for user interface design - opacity of system state, difficulties with error handling, and uncertainty over user intent.
  • the currently underspecified nature of the software architecture for semi- autonomous control (planning and decision making) of the ERAS analogue rover together with the desire to let the complexity of that architecture evolve over time means that the semantic (wordform sequence -> action request mapping) component of the voice interface library must be designed with an eye to modularity and some sense of more or less likely avenues for the coevolution of the voice interface library’s semantic component and the rover’s automated planner/central executive. Crucially, growth in features, especially past the early stage, will involve large additions of code for modest growth of features.

Assumptions and dependencies

Forseeable long-term dependencies:

  • Python 2/3
  • TANGO / PyTango bindings
  • Ubuntu 12.04 LTS
  • CMU Sphinx toolkit

Early/current dependencies:

  • Pocketsphinx [2] and Gstreamer [3], including all of Gstreamer’s dependencies ([4]).
  • Either the Phoenix semantic parser (as maintained by the CMU Olympus project [5]), which is written in C (and for which ctypes Python bindings would have to be written), or the SEMAFOR semantic parser [6-7] (for which Python bindings [8] do exist), which is written in Java. A comparison is available in [9].

Later dependencies:

  • All available open-source spoken dialogue systems that meet basic project interface requirements (see [9]) appear to depend on Java, and further, generally make use of Sphinx 4 [10] rather than Pocketsphinx [2], but it is not clear at this point whether this is good, bad, or if bad, how resolvable.

Probable and future changes in features are described in [1]. Options for spoken dialogue systems are detailed in a review ([9]).

General Constraints

  • ERAS Software Engineering Practices Guidelines [11]
  • The ambient noise level in a helmet is quite high [12]; the CMU Sphinx FAQ [13] offers some suggestions.
  • According to the field testing notes [12], power consumption may be an important issue; Pocketsphinx is the preferred CMU Sphinx library for mobile (e.g. fast, low-battery use) automatic speech recognition, but Sphinx 4 may be an easier option for development.

Objectives

As an interface

  1. Achieve high information efficiency for each voice command. Per reference [14], this measure is chiefly useful as a heuristic for measuring room for improvement (or lack thereof). NB Cruder estimates of this will have to do until a larger database of training data is available.
  2. Appropriate expressivity, primarily for the user (the number of forms the finite set of commands a rover understands can take), but also for feedback. The amount and variety of expressivity allowed will grow as feature requirements 5 and 6 are met more extensively - e.g. for users, more complex grammars and more complex pragmatic understanding on the part of the rover voice agent; for the voice agent, text, synthesized speech, and some amount of supplementary graphical interface will each allow more flexibility in feedback quantity and quality, as situations and users demand.
  3. Learnability is the final interface priority, albeit the lowest one, given difficulty of measurement and that users should have sufficient time to master the capabilities and limitations of the voice interface library, whether in its early, evolving stages or its future, more sophisticated ones.

(See reference [15].)

As software,

  • maximizing ease of feature growth (see requirements 6-8 in [1]) by
    • making the major components (automated speech recognition, spoken language understanding, dialogue management, task management, and natural language generation/feedback) as modular as possible
      • each module should have well-defined and as stable as possible core methods as possible for interfacing modules to call and use;
    • making minimal unwarranted assumptions about the general planning capabilities or architecture of the rover
    • not letting a lightweight system slowly build up more complexity than it can handle, thus making the later move to a spoken dialogue system more painful than it should be: upgrading the spoken dialogue system (particularly its approach to task management/adding new domain-specific agents) to handle generally complex queries and dialogue situations should take priority over (over)extension of a lightweight architecture.
  • meeting the Performance Requirements outlined in [1].

Software Architecture

The core of voice interface library architecture consists of five components:

  1. an automated speech recognition object (responsible for Requirements 1-3)
  2. a spoken language understanding object (responsible for Requirement 4)
  3. a dialogue manager (responsible for Requirements 4, 5)
  4. a task manager (responsible for Requirement 4)
  5. a natural language generation object (the front-end of Requirement 5)

Optionally,

  1. a text-to-speech synthesizer (part of Requirement 6)

may also be a component that should require a minimal amount of additional maintenance or work once it is in place; it would be part of the front-end of Requirement 5, and is part of most (if not all) open source spoken dialogue systems or frameworks.

For the purpose of development, two additional components

  1. a test object
  2. a dummy rover executive - responsible for checking for reasons why a requested action cannot be begun or completed)

are also part of the architecture of the project.

The current plan is for at least component 1 (and possibly also 2) - and if/when it exists, component 6 - to reside on a Tango server onboard an analogue astronaut’s suit, with all other components residing onboard any particular rover, for reasons of minimizing network bandwidth use.

Architecture Development Plan

As mentioned above, the voice interface library is designed with growth in mind. Below is an approximate development ordering, less reliable the further from phase 1 the further forward one goes.

First Phase - ASR and Testing

Implement components 1 and 7.

Draft test scenarios, using use cases in the system requirements document as a guide [1]; record database of plausibly useful commands with different levels of complexity.

Test ASR in Wizard-of-Oz setups (the user’s audio gets piped into the ASR, the ‘person behind the curtain’ sees the resulting recognized text, and relays/requests feedback to the user), asking multiple users to specify as many reasonable variations of plausible commands at each stage as possible (i.e. elicit commands from people unfamiliar with the training corpus). Make few or no assumptions about the rover’s natural language understanding abilities. A simulator would be nice for this; the scenarios mentioned in the prior art research [16] (especially the transcripts studied in Clancey, 2004) may be a good starting place; note also for future reference (1) that ROS comes with a simulation package (2) NASA’s open source Mission Simulation Toolkit [17] may also be of future use here.

In the course of iterative testing (revising the language model and parameters to the ASR model), attempt to optimize all relevant performance requirements - everything except lag-to-feedback and experiment with (or at least record) the amount of training required (or used) for speaker-dependent recognition models.

Second Phase

Focus on component 2 and extend test component appropriately.

The decision must be made whether to (a) start fulfilling (or at least facilitating later fulfillment of) Requirements 6-7 by picking an existing open source dialogue system (which come with components 1-6, and usually Sphinx 4 for component 1) to start integrating (beginning with SLU), or (b) to defer choosing a spoken dialogue system until later and instead start using one of the two identified lightweight spoken language understanding systems (see [9]), with the intention of interfacing it with a later chosen SDS or switching altogether to the native SLU framework and code for the chosen SDS.

Once the SLU system is in place, another round of Wizard-of-Oz tests should resume, documenting and iteratively improving the performance of the SLU component.

Third/Fourth Phases

The next phase should pursue basic dialog management capabilities (component 3) as well as task management and basic feedback (components 4 and 5), with the exact order and degree of detail depending on what seems a more pressing need in test scenarios. Wizard-of-Oz contests should continue as needed. (The test component should continue to expand appropriately.)

Fourth/Fifth Phases

Component 8 should be implemented, requiring further extension of components 5 and 7.

Fifth/Sixth Phases

Requirements 6 and/or 7 can be worked on as needed.