69 lines
4.8 KiB
Python
69 lines
4.8 KiB
Python
|
"""Offline tiny-model heuristic screening of the exact final WAV bytes; never clearance."""
|
||
|
|
from dataclasses import asdict
|
||
|
|
from datetime import datetime, timezone
|
||
|
|
import hashlib
|
||
|
|
from importlib.metadata import version
|
||
|
|
import inspect
|
||
|
|
import json
|
||
|
|
from pathlib import Path
|
||
|
|
import platform
|
||
|
|
import sys
|
||
|
|
import ctranslate2
|
||
|
|
from faster_whisper import WhisperModel
|
||
|
|
|
||
|
|
ROOT = Path(__file__).resolve().parents[3]
|
||
|
|
EVIDENCE = Path(__file__).resolve().parent
|
||
|
|
REVISION = 'd90ca5fe260221311c53c58e660288d3deb8d356'
|
||
|
|
SNAPSHOT = Path.home() / '.cache/huggingface/hub/models--Systran--faster-whisper-tiny/snapshots' / REVISION
|
||
|
|
|
||
|
|
def sha(path):
|
||
|
|
with path.open('rb') as stream: return hashlib.file_digest(stream, 'sha256').hexdigest()
|
||
|
|
|
||
|
|
assert version('faster-whisper') == '1.2.1'
|
||
|
|
started = datetime.now(timezone.utc).isoformat()
|
||
|
|
manifest_path = ROOT / 'public/ambience/manifest.json'
|
||
|
|
manifest_hash = sha(manifest_path)
|
||
|
|
manifest = json.loads(manifest_path.read_text())
|
||
|
|
# Preserve all effective API defaults rather than relying on undocumented invocation settings.
|
||
|
|
parameters = {name: parameter.default for name, parameter in inspect.signature(WhisperModel.transcribe).parameters.items()
|
||
|
|
if name not in ('self', 'audio')}
|
||
|
|
parameters.update(language=None, task='transcribe', beam_size=5, vad_filter=False)
|
||
|
|
ctranslate2.set_random_seed(0)
|
||
|
|
model = WhisperModel(str(SNAPSHOT), device='cpu', compute_type='int8', cpu_threads=4, num_workers=1, local_files_only=True)
|
||
|
|
assert model.model.is_multilingual
|
||
|
|
results = []
|
||
|
|
for entry in manifest['files']:
|
||
|
|
path = ROOT / 'public/ambience' / entry['file']
|
||
|
|
before = sha(path)
|
||
|
|
assert before == entry['sha256'] and path.stat().st_size == entry['bytes']
|
||
|
|
segments, info = model.transcribe(str(path), **parameters)
|
||
|
|
segments = [asdict(segment) for segment in segments] # Exhaust lazy inference; retain exact unedited results.
|
||
|
|
assert sha(path) == before
|
||
|
|
record = {'environmentId': entry['environmentId'], 'file': str(path.relative_to(ROOT)), 'sha256': before,
|
||
|
|
'bytes': path.stat().st_size, 'durationSeconds': entry['duration'],
|
||
|
|
'selectedSegmentStartSeconds': entry['selectedSegmentStartSeconds'], 'sourceSha256': entry['sourceSha256'],
|
||
|
|
'language': info.language, 'languageProbability': info.language_probability,
|
||
|
|
'segments': segments, 'text': ''.join(segment['text'] for segment in segments), 'info': asdict(info)}
|
||
|
|
results.append(record)
|
||
|
|
print(json.dumps({k: record[k] for k in ('environmentId','sha256','language','languageProbability','segments','text')}, ensure_ascii=False), flush=True)
|
||
|
|
assert sha(manifest_path) == manifest_hash
|
||
|
|
required = ['mall', 'market', 'cafe']
|
||
|
|
passed = all(result['segments'] == [] for result in results if result['environmentId'] in required)
|
||
|
|
report = {'schemaVersion': 1, 'startedAtUtc': started, 'completedAtUtc': datetime.now(timezone.utc).isoformat(),
|
||
|
|
'tool': {'name': 'faster-whisper', 'version': version('faster-whisper'), 'python': platform.python_version(),
|
||
|
|
'dependencies': {name: version(name) for name in ('ctranslate2', 'av', 'numpy', 'tokenizers', 'huggingface-hub')}},
|
||
|
|
'model': {'name': 'tiny', 'multilingual': True, 'repository': 'Systran/faster-whisper-tiny', 'revision': REVISION,
|
||
|
|
'files': {path.name: sha(path) for path in sorted(SNAPSHOT.iterdir()) if path.is_file()}},
|
||
|
|
'method': {'input': 'Exact final 12-second public/ambience WAV files, hash verified before and after inference; no source scan, VAD, text editing or manual filtering.',
|
||
|
|
'device': 'cpu', 'computeType': 'int8', 'cpuThreads': 4, 'numWorkers': 1, 'randomSeed': 0,
|
||
|
|
'offline': True, 'parameters': parameters, 'manifestSha256': manifest_hash,
|
||
|
|
'scriptSha256': sha(Path(__file__)), 'command': '/tmp/findword-asr-venv/bin/python .agent/evidence-v0.2.3/privacy-screening/screen-final.py'},
|
||
|
|
'limitations': {'screeningOnly': True, 'humanListeningAcceptance': False, 'legalPrivacyClearance': False,
|
||
|
|
'textIsVerifiedTranscript': False,
|
||
|
|
'explanation': 'Tiny-model ASR is only a screening signal. Empty segments do not prove absence of speech, music or private conversation. Candidate text and language guesses may be hallucinated and are not faithful transcripts. This is NOT human listening acceptance or legal/privacy clearance.'},
|
||
|
|
'requiredEmptyEnvironmentIds': required, 'requiredEmptyPassed': passed, 'files': results}
|
||
|
|
with (EVIDENCE / 'final-asr.json').open('x') as stream:
|
||
|
|
stream.write(json.dumps(report, ensure_ascii=False, indent=2, allow_nan=False) + '\n')
|
||
|
|
if not passed: sys.exit('Required changed excerpts produced ASR segments; screening criterion failed.')
|
||
|
|
print('All three changed excerpts have zero emitted segments. Screening signal only; no listening acceptance or legal/privacy clearance.', flush=True)
|