Heiheihaha17's picture
Add files using upload-large-folder tool
347d9a3 verified
Raw History Blame Contribute Delete
4.57 kB
"""Inspect held-out aligned real speech and static room/scene inventory."""
from __future__ import annotations
import argparse,gzip,json
from collections import Counter
from pathlib import Path
import numpy as np
import soundfile as sf
from dynamic_demo.common import SR,sha,speech_activity,write
def rows(path):
opener=gzip.open if path.suffix=='.gz' else open
with opener(path,'rt') as f:
for line in f:yield json.loads(line)
def source_candidates(release,corpus):
sources={r['sha256']:r for r in rows(corpus/'source_catalog.jsonl')
if r['corpus']=='librispeech' and r['split']=='test'}
alignments={}
for row in rows(release/'alignment_crops.jsonl.gz'):
r=row['crop']
if r['source_sha256'] in sources:alignments.setdefault(r['source_sha256'],r)
candidates=[];file_checks=0
for digest,alignment in sorted(alignments.items()):
source=sources[digest];words=alignment['words'];runs=[];lo=0
for i,w in enumerate(words):
if w['mean_score']<.5:
if i>lo:runs.append((lo,i))
lo=i+1
elif i and w['start_s']-words[i-1]['end_s']>.3:
if i>lo:runs.append((lo,i))
lo=i
if len(words)>lo:runs.append((lo,len(words)))
runs=[(a,b) for a,b in runs if words[b-1]['end_s']-words[a]['start_s']>=3.]
if not runs:continue
# Catalog paths are author-local. Published candidates use corpus-relative paths.
source_path=corpus/'extracted'/Path(source['path']).relative_to('soundspaces_demo/corpora/english_edge_v1/extracted')
if sha(source_path)!=digest:raise ValueError('Original speech checksum mismatch')
wave,sr=sf.read(source_path,dtype='float32')
if sr!=SR or wave.ndim!=1:raise ValueError('Invalid original speech')
file_checks+=1
for first,last in runs:
# VAD runs are recomputed for each final chosen crop; these full-run
# labels provide an early inventory filter only.
a=max(0,round((words[first]['start_s']-.02)*SR))
b=min(len(wave),round((words[last-1]['end_s']+.025)*SR))
spans=speech_activity(wave[a:b],0)
if not spans:continue
longest=max(((y[0]-x[1])/SR for x,y in zip(spans,spans[1:])),default=0.)
candidates.append(dict(source_id=source['source_id'],speaker_id=source['speaker_id'],
source_sha256=digest,original_corpus_path=str(source_path.relative_to(corpus/'extracted')),
original_samples=len(wave),words=words,run_word_indices=[first,last],
aligned_run_duration_s=words[last-1]['end_s']-words[first]['start_s'],
full_run_maximum_internal_silence_s=longest,
aligner=alignment['aligner'],aligner_weights_sha256=alignment['aligner_weights_sha256']))
if file_checks%100==0:print('SOURCE INVENTORY',file_checks,len(candidates),flush=True)
return candidates,file_checks
def inspect(release,corpus,output):
scenes=list(rows(release/'unified_test.jsonl.gz'));rooms=[r for r in rows(release/'rooms.jsonl.gz') if r['split']=='test']
people=[s for s in scenes if s['physical_scene']['humans']>=2]
candidates,files=source_candidates(release,corpus)
report=dict(static_scenes=len(scenes),multi_human_scenes=len(people),test_rooms=len(rooms),
source_files_checked=files,aligned_runs=len(candidates),
source_runs_by_minimum_duration_s={str(t):sum(c['aligned_run_duration_s']>=t and c['full_run_maximum_internal_silence_s']<=.35 for c in candidates) for t in [3.,4.,4.8,5.,6.,7.]},
eligible_parents_by_minimum_duration_s={str(t):dict(scenes=sum(s['frames']/SR>=t for s in people),
rooms=len({s['room_id'] for s in people if s['frames']/SR>=t})) for t in [6.,6.5,7.,7.5,8.]},
source_speakers=sorted({c['speaker_id'] for c in candidates}),
original_test_only=True,minimum_word_alignment_score=.5,maximum_aligned_word_gap_s=.3,
source_catalog_sha256=sha(corpus/'source_catalog.jsonl'),
frozen_release_sha256=sha(release/'release.json'))
write(output/'inventory.json',report);write(output/'source_candidates.json',candidates)
print(json.dumps(report,ensure_ascii=False,indent=2),flush=True)
if __name__=='__main__':
p=argparse.ArgumentParser();p.add_argument('--release',type=Path,required=True);p.add_argument('--corpus',type=Path,required=True);p.add_argument('--output',type=Path,required=True)
a=p.parse_args();inspect(a.release,a.corpus,a.output)