fluency/compute_fluency.py

"""
Compute fluency score from audio file using SRS and PAS calculations
"""

import librosa
import numpy as np
from typing import Dict, Any, Union
from .fluency import calc_srs, calculate_pas, calculate_fluency, get_fluency_insight
from .filler_analyzer import detect_fillers

def compute_fluency_score(file_path: str, whisper_model) -> Dict[str, Any]:
    """
    Compute fluency score and its components from a speech sample.

    Args:
        file_path (str): Path to the audio file.
        whisper_model: Transcription model (e.g., OpenAI Whisper or faster-whisper)

    Returns:
        dict: A dictionary containing fluency score, SRS, PAS, and component scores.
    """
    # Transcribe audio
    result = whisper_model.transcribe(file_path)
    transcript = result.get("text", "").strip()
    segments = result.get("segments", [])

    # Validate early
    if not transcript or not segments:
        raise ValueError("Empty transcript or segments from Whisper.")

    # Detect filler words
    filler_count, _ = detect_fillers(transcript)

    # Load audio
    y, sr = librosa.load(file_path, sr=None)
    duration = len(y) / sr if sr else 0.0
    if duration <= 0:
        raise ValueError("Audio duration invalid or zero.")

    # Calculate pitch variation (in semitones)
    f0, voiced_flags, voiced_probs = librosa.pyin(
        y, sr=sr, fmin=80, fmax=400, frame_length=1024, hop_length=256, fill_na=np.nan)
    voiced_f0 = f0[~np.isnan(f0)]
    pitch_variation = 0.0
    if voiced_f0.size > 0:
        median_f0 = np.nanmedian(voiced_f0)
        median_f0 = max(median_f0, 1e-6)
        semitone_diffs = 12 * np.log2(voiced_f0 / median_f0)
        pitch_variation = float(np.nanstd(semitone_diffs))

    # Analyze pauses
    long_pause_count = 0
    if segments:
        for i in range(len(segments) - 1):
            pause_dur = segments[i + 1]["start"] - segments[i]["end"]
            if pause_dur > 1.0:
                long_pause_count += 1
        # Check beginning and end pauses
        if segments[0]["start"] > 1.0:
            long_pause_count += 1
        if duration - segments[-1]["end"] > 1.0:
            long_pause_count += 1

    # Calculate WPM
    word_count = len(transcript.split())
    words_per_min = (word_count / duration) * 60.0 if duration > 0 else 0.0

    # Calculate SRS - Speech Rate Stability
    srs_score = calc_srs(
        wpm=words_per_min,
        filler_count=filler_count,
        long_pause_count=long_pause_count,
        pitch_variation=pitch_variation
    )

    # Calculate PAS - Pause Appropriateness Score
    pas_result = calculate_pas(
        transcript=transcript,
        segments=segments,
        filler_count=filler_count,
        duration=duration
    )
    pas_score = pas_result["PAS"]

    # Calculate final fluency score
    fluency_result = calculate_fluency(srs=srs_score, pas=pas_score)
    fluency_score = fluency_result["score"]
    insight = get_fluency_insight(fluency_score)

    # Build and return comprehensive result
    return {
        "fluency_score": fluency_score,
        "insight": insight,
        "SRS": srs_score,
        "PAS": pas_score,
        "components": {
            "wpm": words_per_min,
            "filler_count": filler_count,
            "long_pause_count": long_pause_count,
            "pitch_variation": pitch_variation,
            "word_count": word_count,
            "duration": duration,
            "pas_components": pas_result
        },
        "transcript": transcript
    }