Language knowledge beyond acoustics is required for automatic speech recognition.

Evidence

Some information about the language is already embedded in the HMMs of the units of speech, either in the form of word models or as the phone inventory. However, an extensive knowledge of language must be embedded in the recognizer. Most commonly, the grammatical constraints are learned as statistical dependences in the context of words from large text databases (Goodman, 2001). When the vocabulary is small, it may be feasible to make a list of allowed words and define their HMMs. However, when the size of the vocabulary is larger, acoustically similar words are hard to distinguish during recognition. Detecting boundaries of words in continuous speech is also difficult, which adds to the challenge. It becomes necessary to utilize either task-specific or general constraints to overcome these problems. — Ville Pulkki, Matti Karjalainen, Communication Acoustics- An…, 2274

Categorization

• DEFINE: Explains why grammar and vocabulary knowledge must be layered atop HMM phone models.
• QUESTION: What specific statistical language-modeling techniques handle acoustically similar words in large vocabularies?