Mid-1970s, Carnegie Mellon: the Hearsay-II project is trying to make a computer understand continuous speech — a ~1,000-word vocabulary, real sentences, noisy signal. The knowledge that bears on the problem is wildly heterogeneous: acoustics, phonetics, syllable structure, a lexicon, syntax, semantics of the domain. Every piece is partial and errorful — the acoustics can’t reliably tell you the phoneme, the lexicon can’t be sure where words begin, the grammar can rule things out but not hear anything.
The obvious architecture is a pipeline: signal → phonemes → words → parse → meaning, each stage consuming the previous stage’s output. Every speech system before Hearsay-II more or less tried this. It failed, and the question above asks you to identify the architectural reason.
What the Hearsay team built instead defined a new architecture, the blackboard system:
- The blackboard: one shared data structure holding hypotheses at every level of abstraction (this stretch of signal might be an “S”; these syllables might be the word “seven”; this phrase might be a database query) — each with a confidence, each linked to the hypotheses that support it.
- Knowledge sources (KSs): independent specialist modules — one knows acoustics, one knows the lexicon, one knows grammar. Each is condition/action, like a production grown to expert size: when hypotheses of a certain kind appear, then I can create or re-rate others. KSs never call each other; they communicate only through the board.
- Control: a scheduler that watches what changed and decides which eager KS runs next — opportunistically, working wherever the board currently looks most promising, rather than in any fixed order.
The metaphor the team used: specialists standing around a physical blackboard, each stepping up to write when — and only when — their expertise applies to what’s currently written.
So: why did speech demand this? Pick the property that killed the pipeline.