""" Compare VibeVoice vs Fish S2 Pro vs Sooktam-2 on Modi Hindi voice. Same text, same reference audio, measure quality + speed. """ import os import time import torch import soundfile as sf import numpy as np os.makedirs("/home/ubuntu/comparison_samples", exist_ok=True) REF_AUDIO = "/home/ubuntu/vibevoice/demo/voices/modi.wav" REF_TEXT = "मेरे प्यारे देशवासियों, मुझे सीतापुर के ओजस्वी ने लिखा है कि अमृत महोत्सव से जुड़ी चर्चाएं उन्हें खूब पसंद आ रही हैं।" TEXTS = { "medium": "मेरे प्यारे देशवासियों, आज मैं आपके साथ कुछ बहुत ज़रूरी बातें करना चाहता हूँ. हमारा देश एक नये दौर में प्रवेश कर रहा है.", "speech": "भारत आज दुनिया की पाँचवीं सबसे बड़ी अर्थव्यवस्था है. हमारे युवाओं की ऊर्जा, हमारे वैज्ञानिकों की प्रतिभा, और हमारे किसानों की मेहनत, यही हमारी असली ताकत है.", } def test_fish_s2(): print("\n" + "=" * 60) print("FISH AUDIO S2 PRO") print("=" * 60) try: from fish_speech.inference import load_model, inference # Try the HF approach from huggingface_hub import hf_hub_download import subprocess print("Downloading Fish S2 Pro model...") # Use fish_speech CLI approach for label, text in TEXTS.items(): t0 = time.perf_counter() result = subprocess.run([ "python3", "-m", "fish_speech.inference", "--model", "fishaudio/s2-pro", "--text", text, "--reference", REF_AUDIO, "--output", f"/home/ubuntu/comparison_samples/fish_{label}.wav", ], capture_output=True, text=True, timeout=120) gen = time.perf_counter() - t0 if os.path.exists(f"/home/ubuntu/comparison_samples/fish_{label}.wav"): data, sr = sf.read(f"/home/ubuntu/comparison_samples/fish_{label}.wav") dur = len(data) / sr print(f" {label}: {dur:.2f}s audio | {gen:.2f}s gen | RTF={gen/dur:.3f}x") else: print(f" {label}: FAILED - {result.stderr[:200]}") except Exception as e: print(f" Fish S2 Pro failed: {e}") print(" Trying alternative approach...") try: from fish_speech.models.text2semantic.inference import launch as fish_launch print(" fish_speech models available, but need setup") except Exception as e2: print(f" Alternative also failed: {e2}") def test_sooktam2(): print("\n" + "=" * 60) print("SOOKTAM-2 (BharatGen)") print("=" * 60) try: import sys sys.path.insert(0, "/home/ubuntu/sooktam2") # Check if setup-cls.sh needs to run if os.path.exists("/home/ubuntu/sooktam2/setup-cls.sh"): print("Running setup-cls.sh...") import subprocess result = subprocess.run(["bash", "setup-cls.sh"], cwd="/home/ubuntu/sooktam2", capture_output=True, text=True, timeout=120) if result.returncode != 0: print(f" setup-cls.sh stderr: {result.stderr[:300]}") from transformers import AutoModel print("Loading Sooktam-2 model...") model = AutoModel.from_pretrained( "/home/ubuntu/sooktam2", trust_remote_code=True, ) print(f" Loaded. VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB") for label, text in TEXTS.items(): t0 = time.perf_counter() wav, sr, _ = model.infer( ref_file=REF_AUDIO, ref_text=REF_TEXT, gen_text=text, tokenizer="cls", cls_language="hindi", file_wave=f"/home/ubuntu/comparison_samples/sooktam_{label}.wav", ) gen = time.perf_counter() - t0 dur = len(wav) / sr rtf = gen / dur if dur > 0 else 999 print(f" {label}: {dur:.2f}s audio | {gen:.2f}s gen | RTF={rtf:.3f}x | VRAM={torch.cuda.memory_allocated()/1e9:.1f}GB") del model torch.cuda.empty_cache() except Exception as e: print(f" Sooktam-2 failed: {e}") import traceback traceback.print_exc() def test_vibevoice(): print("\n" + "=" * 60) print("VIBEVOICE 1.5B (our current best)") print("=" * 60) try: from vibevoice.modular.modeling_vibevoice_inference import VibeVoiceForConditionalGenerationInference from vibevoice.processor.vibevoice_processor import VibeVoiceProcessor processor = VibeVoiceProcessor.from_pretrained("microsoft/VibeVoice-1.5B") try: model = VibeVoiceForConditionalGenerationInference.from_pretrained( "microsoft/VibeVoice-1.5B", torch_dtype=torch.bfloat16, device_map="cuda", attn_implementation="flash_attention_2") except: model = VibeVoiceForConditionalGenerationInference.from_pretrained( "microsoft/VibeVoice-1.5B", torch_dtype=torch.bfloat16, device_map="cuda", attn_implementation="sdpa") model.eval() model.set_ddpm_inference_steps(num_steps=10) # Warmup inp = processor(text=["Speaker 1: test."], voice_samples=[[REF_AUDIO]], padding=True, return_tensors="pt", return_attention_mask=True) for k, v in inp.items(): if torch.is_tensor(v): inp[k] = v.to("cuda") _ = model.generate(**inp, max_new_tokens=None, cfg_scale=1.3, tokenizer=processor.tokenizer, generation_config={"do_sample": False}, verbose=False, is_prefill=True, show_progress_bar=False) for label, text in TEXTS.items(): full_text = f"Speaker 1: {text}" inp = processor(text=[full_text], voice_samples=[[REF_AUDIO]], padding=True, return_tensors="pt", return_attention_mask=True) for k, v in inp.items(): if torch.is_tensor(v): inp[k] = v.to("cuda") torch.manual_seed(42) torch.cuda.synchronize() t0 = time.perf_counter() out = model.generate(**inp, max_new_tokens=None, cfg_scale=1.3, tokenizer=processor.tokenizer, generation_config={"do_sample": False}, verbose=False, is_prefill=True, show_progress_bar=False) torch.cuda.synchronize() gen = time.perf_counter() - t0 if out.speech_outputs[0] is not None: path = f"/home/ubuntu/comparison_samples/vibevoice_{label}.wav" processor.save_audio(out.speech_outputs[0], output_path=path) dur = out.speech_outputs[0].shape[-1] / 24000 print(f" {label}: {dur:.2f}s audio | {gen:.2f}s gen | RTF={gen/dur:.3f}x | VRAM={torch.cuda.memory_allocated()/1e9:.1f}GB") del model torch.cuda.empty_cache() except Exception as e: print(f" VibeVoice failed: {e}") import traceback traceback.print_exc() if __name__ == "__main__": print("COMPARISON: VibeVoice vs Fish S2 Pro vs Sooktam-2") print(f"Reference audio: {REF_AUDIO}") print(f"Language: Hindi") test_sooktam2() torch.cuda.empty_cache() test_vibevoice() torch.cuda.empty_cache() test_fish_s2() print("\n" + "=" * 60) print("All samples saved in /home/ubuntu/comparison_samples/") print("=" * 60)