Upload folder using huggingface_hub

dataset/mainstream_us_english/female/hesitation/74-mainstream_us_english-female-1388-hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4959261b84a8d979249ee25f1285d99edf306c7bedad421448e444865884e204
+size 998444

dataset/mainstream_us_english/female/hesitation/75-mainstream_us_english-female-1388-hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ad2969f77c741c55f0f992b93791e779ef668760f7ae4225fcb33bbf4aaf22ea
+size 925484

dataset/mainstream_us_english/female/hesitation/76-mainstream_us_english-female-1388-hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2a11adc2cd93e142d6e9d6c72d119b3e106f1b509eecd6bd4121cc3f3b8100d
+size 890924

dataset/mainstream_us_english/female/hesitation/77-mainstream_us_english-female-1388-hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5705267134e02b684e7c07c7d459a7821991a47862faa787b3ebff50e9148ab1
+size 1006124

dataset/mainstream_us_english/female/hesitation/78-mainstream_us_english-female-1388-hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddb4e17aea2593ed1641bfec9a932913a75c02c5c9398ea7f8c7553e016ee7d9
+size 937004

dataset/mainstream_us_english/female/hesitation/79-mainstream_us_english-female-1388-hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a1d870b6b2f10fd2465025bdc8001523841962db5675c089d05eb2c319732a05
+size 890924

dataset/mainstream_us_english/female/no_hesitation/0-mainstream_us_english-female-1388-no_hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:312c2fcf155098f2f31ffd413f7db5884dcd7273d4cd23d4a74817e1be62ad82
+size 802604

dataset/mainstream_us_english/female/no_hesitation/1-mainstream_us_english-female-1388-no_hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ab7789cada01007aed17cd45aecd930b090cb2d2c752aeb25c7d236cad3070d
+size 825644

dataset/mainstream_us_english/female/no_hesitation/2-mainstream_us_english-female-1388-no_hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77ad60c05ae8e8a907ff2cf678ce44d52372bcc9c83b27b85b6fd33cbbefc25c
+size 725804

dataset/mainstream_us_english/female/no_hesitation/3-mainstream_us_english-female-1388-no_hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a97e614a274beb8a93ff5d60fac9c93c0a1e6e4392d6786ae11f06eef84f4c6c
+size 679724

dataset/mainstream_us_english/female/no_hesitation/4-mainstream_us_english-female-1388-no_hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f4447a749576352978edeedb0b98cf9f3831e365bade1ce0a8bca8bcb547462
+size 706604

dataset/mainstream_us_english/female/no_hesitation/5-mainstream_us_english-female-1388-no_hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7c4ffcdee198034b2425c21d0421b397d252ceec5ceaf9562fe166ddbf7843a6
+size 963884

dataset/mainstream_us_english/female/no_hesitation/6-mainstream_us_english-female-1388-no_hesitation.wav

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9caa3c801ea7cbecbe833f2d2717b5581dbb5fef71e07be434f8d465b3cad1a1
+size 860204

generate_speech.py

@@ -0,0 +1,279 @@
+#!/usr/bin/env python3
+"""
+Generate speech samples using MegaTTS3 for different accents and genders.
+
+This script:
+1. Loads question metadata from JSON files
+2. Filters and constructs accent/gender dictionaries
+3. Generates speech using MegaTTS3 voice cloning
+"""
+
+import json
+import shutil
+import time
+from pathlib import Path
+
+import pandas as pd
+from gradio_client import Client, handle_file
+from tqdm.auto import tqdm
+
+
+def generate_with_retry(client, speaker_audio_path, text, max_retries=5, initial_delay=1):
+    """Generate speech with retry logic and exponential backoff.
+    
+    Args:
+        client: Gradio client instance
+        speaker_audio_path: Path to reference speaker audio
+        text: Text to synthesize
+        max_retries: Maximum number of retry attempts (default: 5)
+        initial_delay: Initial delay in seconds (default: 1)
+        
+    Returns:
+        Path to generated audio file
+        
+    Raises:
+        Exception: If all retry attempts fail
+    """
+    for attempt in range(max_retries):
+        try:
+            result = client.predict(
+                inp_audio=handle_file(speaker_audio_path),
+                inp_text=text,
+                infer_timestep=32,
+                p_w=1.4,
+                t_w=3,
+                api_name="/generate_speech"
+            )
+            return result
+        except Exception as e:
+            if attempt == max_retries - 1:
+                # Last attempt failed, raise the exception
+                raise
+            # Calculate backoff delay: 1s, 2s, 4s, 8s, 16s
+            delay = initial_delay * (2 ** attempt)
+            print(f"\nRetry {attempt + 1}/{max_retries} after error: {str(e)[:100]}")
+            print(f"Waiting {delay}s before retrying...")
+            time.sleep(delay)
+
+
+def build_accent_gender_dicts(questions_dir="questions"):
+    """Build accent-gender dictionaries from question JSON files.
+    
+    Args:
+        questions_dir: Directory containing question JSON files
+        
+    Returns:
+        tuple: (accents_gender, accent_gender_both, final_dict)
+    """
+    questions_path = Path(questions_dir)
+    
+    # Step 1: Build initial accents_gender dictionary
+    print("Building accent-gender mappings...")
+    accents_gender = {}
+    for item in sorted(questions_path.glob("*.json")):
+        with open(item, "r") as f:
+            data = json.load(f)
+        
+        # Filter based on text length and other criteria
+        if len(data["text"]) > 50:
+            # Skip certain US English speakers
+            if data["accent"] == "Mainstream US English" and "US English" not in data["l1"]:
+                continue
+            # Skip specific problematic speaker
+            if data["speaker"] == "EDACC-C19-B":
+                print(f"Skipping speaker {data['speaker']} - {item.stem}")
+                continue
+            
+            accent_gender_key = f"{data['accent']}-{data['gender']}"
+            if accent_gender_key not in accents_gender:
+                accents_gender[accent_gender_key] = item.stem
+    
+    print(f"Found {len(accents_gender)} accent-gender combinations")
+    
+    # Step 2: Filter to only include accent-gender pairs where both male and female exist
+    print("\nFiltering for balanced gender representation...")
+    accent_gender_both = {}
+    for k in accents_gender:
+        if "-male" in k:
+            if k.replace("-male", "-female") in accents_gender:
+                accent_gender_both[k] = accents_gender[k]
+            else:
+                print(f"Missing female pair for: {k}")
+        elif "-female" in k:
+            if k.replace("-female", "-male") in accents_gender:
+                accent_gender_both[k] = accents_gender[k]
+            else:
+                print(f"Missing male pair for: {k}")
+    
+    print(f"Balanced pairs: {len(accent_gender_both)}")
+    
+    # Show unique accents
+    unique_accents = set([x.split("-")[0] for x in accent_gender_both.keys()])
+    print(f"\nUnique accents: {sorted(unique_accents)}")
+    
+    # Step 3: Filter to specific accents of interest
+    keys_to_keep = [
+        "Mainstream US",
+        "Southern British",
+        "Indian",
+        "Chinese",
+        "Latin American",
+        "Eastern European",
+    ]
+    
+    final_dict = {}
+    for k in accent_gender_both:
+        if any(x in k for x in keys_to_keep):
+            final_dict[k] = accent_gender_both[k]
+    
+    print(f"\nFinal selection: {len(final_dict)} accent-gender pairs")
+    for k, v in sorted(final_dict.items()):
+        print(f"  {k}: {v}")
+    
+    return accents_gender, accent_gender_both, final_dict
+
+
+def generate_speech_samples(
+    final_dict,
+    prompts_csv="tts_prompts.csv",
+    output_dir="dataset",
+    questions_dir="questions",
+    gradio_endpoint="cdminix/MegaTTS3-Voice-Cloning"
+):
+    """Generate speech samples for all accent-gender pairs.
+    
+    Args:
+        final_dict: Dictionary mapping accent-gender keys to question IDs
+        prompts_csv: Path to CSV file containing prompts
+        output_dir: Base directory for output files
+        questions_dir: Directory containing reference speaker audio
+        gradio_endpoint: Gradio API endpoint for MegaTTS3
+    """
+    # Initialize MegaTTS3 client
+    print(f"\nInitializing MegaTTS3 client: {gradio_endpoint}")
+    client = Client(gradio_endpoint)
+    
+    # Load prompts
+    df = pd.read_csv(prompts_csv)
+    df_hesitation = df[df["has_disfluency"] == True]
+    df_no_hesitation = df[df["has_disfluency"] == False]
+    
+    print(f"\nTotal prompts: {len(df)} ({len(df_hesitation)} with hesitation, {len(df_no_hesitation)} without)")
+    
+    result_rows = []
+    
+    # Calculate total number of files to generate
+    total_files = len(final_dict.keys()) * (len(df_hesitation) + len(df_no_hesitation))
+    print(f"Total files to generate: {total_files}")
+    
+    pbar = tqdm(total=total_files, desc="Generating audio")
+    
+    for k in sorted(final_dict.keys()):
+        accent = k.split("-")[0]
+        accent = accent.replace(" ", "_").lower()
+        gender = k.split("-")[1]
+        
+        # Create output directories
+        hesitation_dir = Path(output_dir) / accent / gender / "hesitation"
+        no_hesitation_dir = Path(output_dir) / accent / gender / "no_hesitation"
+        hesitation_dir.mkdir(parents=True, exist_ok=True)
+        no_hesitation_dir.mkdir(parents=True, exist_ok=True)
+        
+        # Get speaker audio path
+        speaker_audio_path = Path(questions_dir) / f"{final_dict[k]}.wav"
+        speaker_audio_path_corrected = Path(questions_dir) / f"{final_dict[k]}_corrected.wav"
+        
+        # Use corrected version if available, otherwise copy original
+        if not speaker_audio_path_corrected.exists():
+            if speaker_audio_path.exists():
+                shutil.copy(speaker_audio_path, speaker_audio_path_corrected)
+        
+        speaker_audio_path = str(speaker_audio_path_corrected)
+        
+        # Generate hesitation samples
+        for idx, row in df_hesitation.iterrows():
+            _id = f"{idx}-{accent}-{gender}-{final_dict[k]}-hesitation"
+            output_path = hesitation_dir / f"{_id}.wav"
+            
+            if output_path.exists():
+                result_rows.append({
+                    "id": _id,
+                    "accent": accent,
+                    "gender": gender,
+                    "speaker_audio": speaker_audio_path,
+                })
+                pbar.update(1)
+                continue
+            
+            # Generate with retry logic
+            result = generate_with_retry(client, speaker_audio_path, row["text"])
+            
+            # Copy the result file to output path
+            shutil.copy(result, output_path)
+            
+            result_rows.append({
+                "id": _id,
+                "accent": accent,
+                "gender": gender,
+                "speaker_audio": speaker_audio_path,
+            })
+            pbar.update(1)
+        
+        # Generate no-hesitation samples
+        for idx, row in df_no_hesitation.iterrows():
+            _id = f"{idx}-{accent}-{gender}-{final_dict[k]}-no_hesitation"
+            output_path = no_hesitation_dir / f"{_id}.wav"
+            
+            if output_path.exists():
+                result_rows.append({
+                    "id": _id,
+                    "accent": accent,
+                    "gender": gender,
+                    "speaker_audio": speaker_audio_path,
+                })
+                pbar.update(1)
+                continue
+            
+            # Generate with retry logic
+            result = generate_with_retry(client, speaker_audio_path, row["text"])
+            
+            # Copy the result file to output path
+            shutil.copy(result, output_path)
+            
+            result_rows.append({
+                "id": _id,
+                "accent": accent,
+                "gender": gender,
+                "speaker_audio": speaker_audio_path,
+            })
+            pbar.update(1)
+    
+    pbar.close()
+    
+    # Save results
+    df_result = pd.DataFrame(result_rows)
+    df_result.to_csv("tts_prompts_result.csv", index=False)
+    print(f"\nCompleted! Generated {len(result_rows)} audio files.")
+    print(f"Results saved to: tts_prompts_result.csv")
+
+
+def main():
+    """Main execution function."""
+    print("=" * 70)
+    print("Speech Generation Pipeline")
+    print("=" * 70)
+    
+    # Step 1: Build dictionaries
+    accents_gender, accent_gender_both, final_dict = build_accent_gender_dicts()
+    
+    # Step 2: Generate speech samples
+    generate_speech_samples(final_dict)
+    
+    print("\n" + "=" * 70)
+    print("Pipeline complete!")
+    print("=" * 70)
+
+
+if __name__ == "__main__":
+    main()
+

requirements.txt

@@ -7,4 +7,7 @@ torch
 datasets[audio]
 torchcodec
 python-dotenv
-replicate
+replicate
+gradio_client
+transformers
+pandas