Speech-to-Text & Normalization

Offline voice capture relies on the Vosk Speech-to-Text library running local acoustic models, paired with a custom number normalization algorithm that converts spelled-out numbers into standard numeric arguments.

1. Audio Recording Specification

Format: 16,000Hz Sample Rate, 16-bit Mono, Linear PCM WAV format.
Buffer Size: Recorded in chunks on-device, then uploaded as raw WAV files to the backend chatbot router.

2. Word-to-Number Normalization

Speech recognizers generally transcribe numeric arguments as literal words (e.g. "walk forward five steps"). To allow the parser to extract precise integers for gait loops, the backend employs a recursive word-number solver:

def convert_words_to_numbers(text):
    # Splits, checks against value dictionary, handles multipliers:
    # "two hundred and forty five" -> 245
    #
    # 1. Matches scale terms: "hundred", "thousand"
    # 2. Accumulates totals: tens + units
    # 3. Swaps words dynamically within the sentence text

Voice Command Data Flow

Vosk STT and Semantic Parsing Data Flow