| """Validate continuation, emitted PCM and shared scoring; no subjective claims.""" |
| import json |
| import numpy as np |
| import torch |
| import torchaudio |
|
|
| import settings |
| from engine import SmallEngine, Scorers |
| from beam_search import ChunkSearch |
| from prompting import build |
|
|
|
|
| class FixedScorer: |
| device='none (constant validation scorer)' |
| last_timing={'asr_seconds':0.,'quality_seconds':0.,'total_seconds':0.} |
| def score(self,waves,text,final): |
| return [{'reward':1.,'prefix_wer':0.,'hyp':'','sec':len(w)/48000} for w in waves] |
|
|
|
|
| def main(): |
| engine=SmallEngine() |
| spec=build({'text':'You knew what this meant to me. Tell me the truth.', |
| 'caption':'Intense contained anger, clear words.'},'caption','en',seconds=6.6) |
| reference=settings.HERE/'performance/benchmark-reference.wav' |
| baseline=engine.generate_many([spec]*4,1.5,seed=991,reference_path=reference,streaming=True, |
| batch_size=4,chunk_frames=10,max_frames=20) |
| codes=engine.last_codes[0].clone() |
| emitted=[] |
| wave,p=ChunkSearch(engine,FixedScorer()).generate(spec,n=4,beam_width=1,chunk_frames=10, |
| max_seconds=1.6,seed=991,reference_path=reference,on_commit=lambda w,e:emitted.append(w)) |
| new=engine.last_codes[0] |
| continuation={'same_tokens':bool(torch.equal(codes,new)), |
| 'token_agreement':float((codes[:min(len(codes),len(new))]==new[:min(len(codes),len(new))]).float().mean()), |
| 'same_lengths':len(baseline[0])==len(wave), |
| 'pcm_rms_difference':float(np.sqrt(np.mean((baseline[0][:len(wave)]-wave)**2))) if len(baseline[0])>=len(wave) else None, |
| 'commits_equal_final':bool(np.array_equal(np.concatenate(emitted),wave)), |
| 'chunk_count':len(emitted)} |
| assert continuation['same_tokens'] and continuation['commits_equal_final'],continuation |
| scorers=Scorers() |
| |
| |
| from transformers import AutoModel |
| gen_tower=AutoModel.from_pretrained('/mnt/nvme/moss-15-v2-assets/vc_genuineness/voiceclap_commercial', |
| trust_remote_code=True).to(scorers.retriever.device).eval() |
| blend_tower=AutoModel.from_pretrained('/mnt/nvme/moss-15-v2-assets/vc_blend/voiceclap_commercial', |
| trust_remote_code=True).to(scorers.retriever.device).eval() |
| audio=torchaudio.functional.resample(torch.tensor(wave,device=scorers.retriever.device),48000,16000)[None] |
| with torch.inference_mode(): |
| reference_embedding=scorers.retriever.model.encode_waveform(audio).float() |
| gen_embedding=gen_tower.encode_waveform(audio).float() |
| blend_embedding=blend_tower.encode_waveform(audio).float() |
| shared={'genuineness_embedding_max_abs':float((reference_embedding-gen_embedding).abs().max()), |
| 'blend_embedding_max_abs':float((reference_embedding-blend_embedding).abs().max())} |
| assert shared['genuineness_embedding_max_abs']<1e-6 and shared['blend_embedding_max_abs']<1e-6,shared |
| result={'continuation':continuation,'shared_audio_tower':shared, |
| 'notes':'Numerical/causal validation on one prompt, four candidates, twenty frames; not a human perceptual-equivalence study.'} |
| (settings.HERE/'search/validation.json').write_text(json.dumps(result,indent=2)) |
| print(json.dumps(result),flush=True) |
|
|
|
|
| if __name__=='__main__':main() |
|
|
