README.md

metadata

title: Fluency Benchmark
emoji: 🎙️
colorFrom: blue
colorTo: green
sdk: docker
app_port: 7860
app_file: app.py
pinned: false

Fluency Benchmark App

Upload an English speech audio file to measure temporal fluency — flow, continuity, and pausing behavior.

What It Measures

• Temporal fluency — flow and continuity of speech
• Pause behavior — frequency, duration, and placement of pauses
• Articulation — smoothness of delivery (legato vs staccato)
• Hesitation diagnosis — cognitive load and utterance constraints

What it does NOT measure: grammar, vocabulary, pronunciation, or accent.

Pipeline Stages

1. VAD (Silero) — detects speech vs silence, computes 6 temporal features
2. Transcription (WhisperX base) — word-level aligned transcript
3. Placement — classifies each pause as boundary-aligned or mid-clause
4. FA Features — word duration, confidence, filled pauses, speech rate CV
5. Syntactic — POS-tags pauses as before content vs function words
6. Inference — 6 ordinal models + 1 dominance model → predictions
7. Composite — 6 dimensions weighted → percentile + confidence interval

Output

• Fluency Percentile (0-100, relative to benchmark population of 917 speakers)
• Fluency Band (LOW / MEDIUM / HIGH)
• 6 Dimension Scores (Continuity, Pause Quality, Placement, Articulation, Dominance, Word Precision)
• Ordinal Predictions (Pause Frequency, Duration, Placement, Cognitive Load, Utterance Constraints, Articulation)
• 95% Confidence Interval (Dirichlet bootstrap)