import os
import warnings

import whisper
import librosa
import numpy as np
import tensorflow as tf
import tensorflow_hub as hub
import csv

# Reduce TensorFlow log noise and avoid attempting GPU / oneDNN on CPU-only envs.
# NOTE: These env vars must be set before TensorFlow fully initializes; setting them
# here greatly reduces, but may not completely remove, startup logs on some platforms.
os.environ.setdefault("TF_CPP_MIN_LOG_LEVEL", "2")  # hide INFO/WARNING logs
os.environ.setdefault("TF_ENABLE_ONEDNN_OPTS", "0")  # disable oneDNN custom ops
os.environ.setdefault("CUDA_VISIBLE_DEVICES", "-1")  # don't try to use CUDA GPUs

# Suppress specific library warnings that are expected in this setup.
warnings.filterwarnings(
    "ignore",
    category=UserWarning,
    message="FP16 is not supported on CPU; using FP32 instead",
)
warnings.filterwarnings(
    "ignore",
    category=FutureWarning,
    module="librosa",
)

# Load ASR
asr_model = whisper.load_model("small")

# Load YAMNet for sound classification
yamnet = hub.load("https://tfhub.dev/google/yamnet/1")
class_map_path = yamnet.class_map_path().numpy()
if isinstance(class_map_path, bytes):
    class_map_path = class_map_path.decode("utf-8")

# Parse YAMNet class map CSV to get human-readable labels
with tf.io.gfile.GFile(class_map_path) as f:
    reader = csv.DictReader(f)
    yamnet_labels = [row["display_name"] for row in reader]

# Simple Emotion Estimator (from YAMNet embedding)
def estimate_emotion(activation):
    mean_val = activation.mean()
    if mean_val > 0.3:
        return "Happy / Excited"
    elif mean_val < -0.3:
        return "Sad / Depressed"
    return "Neutral"


def speech_to_text(audio):
    # Force FP32 on CPU to avoid FP16 warnings and ensure compatibility
    result = asr_model.transcribe(audio, fp16=False)
    return result["text"]


def detect_sound(audio):
    # Load mono waveform at 16 kHz as 1D float32 array, as expected by YAMNet
    waveform, sr = librosa.load(audio, sr=16000, mono=True)
    waveform = waveform.astype(np.float32)
    scores, embeddings, _ = yamnet(waveform)
    mean_scores = np.mean(scores.numpy(), axis=0)
    top_idx = int(np.argmax(mean_scores))
    # Look up human-readable class label from YAMNet's CSV class map
    label = yamnet_labels[top_idx] if 0 <= top_idx < len(yamnet_labels) else "Unknown"
    return label, float(mean_scores.max())


def analyze_audio(audio_file):
    summary = {}

    summary["transcription"] = speech_to_text(audio_file)

    event, confidence = detect_sound(audio_file)
    summary["sound_event"] = event
    summary["sound_confidence"] = float(confidence)

    summary["emotion"] = "Neutral (approx)"

    summary["speakers"] = "Not available in HF-free version"

    return summary