src/detector.py

import os
from huggingface_hub import snapshot_download
from transformers import pipeline

# Use /tmp which is always writable in Hugging Face Spaces
cache_dir = "/tmp/hf_cache"
os.makedirs(cache_dir, exist_ok=True)
os.environ["TRANSFORMERS_CACHE"] = cache_dir

# Download model to the writable tmp directory
model_path = snapshot_download(repo_id="openai/whisper-base", cache_dir=cache_dir)

# Load the pipeline from the local path
asr = pipeline("automatic-speech-recognition", model=model_path)

def detect_accent(audio_path: str):
    result = asr(audio_path, return_timestamps=True)
    text = result["text"].lower()

    # Score dictionary for multiple accents
    accent_scores = {
        "American": 0,
        "British": 0,
        "Australian": 0,
        "Indian": 0,
    }

    # American patterns
    if any(word in text for word in ["gonna", "wanna", "dude", "gotta"]):
        accent_scores["American"] += 2
    if any(word in text for word in ["elevator", "sidewalk", "apartment"]):
        accent_scores["American"] += 1

    # British patterns
    if any(word in text for word in ["mate", "cheers", "lorry", "flat", "rubbish"]):
        accent_scores["British"] += 2
    if any(word in text for word in ["colour", "favourite", "centre"]):
        accent_scores["British"] += 1

    # Australian patterns
    if any(word in text for word in ["yeah nah", "arvo", "barbie", "brekkie"]):
        accent_scores["Australian"] += 2
    if "g’day" in text or "no worries" in text:
        accent_scores["Australian"] += 1

    # Indian English patterns
    if any(phrase in text for phrase in ["kindly do the needful", "revert back", "only", "prepone"]):
        accent_scores["Indian"] += 2
    if any(word in text for word in ["co-brother", "timepass", "out of station"]):
        accent_scores["Indian"] += 1

    # Determine best guess or fallback to "Unknown"
    top_accent = max(accent_scores, key=accent_scores.get)
    if accent_scores[top_accent] == 0:
        return "Unknown", 50, text  # default fallback

    confidence = accent_scores[top_accent] * 10 + 50  # simple mock scoring
    return top_accent, min(confidence, 95), text