Introduction
Astronauts on the International Space Station (ISS) operate in a demanding microgravity environment where traditional methods of following complex procedures—such as paper manuals or handheld tablets—are impractical. If a manual is not secured, it floats away; if an astronaut is holding a device, they lose the use of a hand critical for maintenance or scientific tasks. Clarissa was developed as the first spoken dialogue system deployed on the ISS to address these challenges, providing a hands-free, eyes-free "procedure navigator."
Technical Architecture
The Clarissa system relies on three core technical pillars to ensure reliability in a high-stakes environment:
- Grammar-Based Recognition: Utilizing the Regulus toolkit, the system extracts language models from a large unification grammar. This approach was chosen over statistical models due to the scarcity of initial training data and the preference of astronaut crews for the "controlled language" structures common in aviation and military contexts.
- Open Microphone Discrimination: A significant challenge in the ISS environment is "open microphone" noise—distinguishing between commands intended for the system and general ambient chatter or conversations between crew members. Clarissa employs Support Vector Machines (SVMs) to classify audio input.
- Side-Effect Free Dialogue Management: To ensure safety, the system treats dialogue states as objects with "back-pointers." This allows for a robust "undo" capability, enabling users to revert the system state easily if a mistake occurs or a command is misinterpreted.
Evaluation and Performance
The system's performance was rigorously evaluated using a library of XML-encoded procedures and a large corpus of over 10,000 recorded utterances. A key finding was that grammar-based models (GLM) significantly outperformed statistical language models (SLM) in this niche application. While Word Error Rates were comparable, the semantic error rate for GLM was much lower (6.3% vs. 10.2%), as the grammar-based approach better aligns with the underlying logic of technical procedures.
The most critical advancement was the use of a Quadratic SVM for input rejection. By analyzing both recognizer confidence and lexical features (the specific words used), the system drastically reduced task-level errors.
As shown in the performance data, the configuration combining Grammar-based recognition with a Quadratic SVM achieved the lowest overall task error rate of 5.4%. This setup was particularly effective at rejecting "out-of-domain" speech, which is essential for preventing the system from acting on background conversations.
Conclusion and Future Outlook
Clarissa was successfully deployed to the ISS on January 14, 2005. Its deployment proved that a voice-enabled interface could effectively manage complex procedures in space. However, challenges remain in scaling the system. Currently, procedures must be manually XML-encoded, a labor-intensive process. Future work will focus on automating this pipeline and further hardening the system against the unique acoustic environment of the ISS, including persistent fan noise and echoes.
Summary updated