File size: 3,287 Bytes
b21cfae | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 | """Actual loaded TYGER checks, never another model load; synthetic audio only."""
import asyncio
from pathlib import Path
import subprocess
import time
import numpy as np
def read16(path,seconds=6):
p=subprocess.run(['ffmpeg','-nostdin','-v','error','-i',str(path),'-t',str(seconds),
'-ar','16000','-ac','1','-f','f32le','pipe:1'],capture_output=True,check=True)
x=np.frombuffer(p.stdout,'<f4').copy()
if not len(x):raise ValueError('Empty synthetic fixture')
return x
async def model_checks(ears):
root=Path(__file__).resolve().parents[1]
# A sustained reply, not a five-second clip followed by three seconds of
# "speaking" silence (which tests a different turn transition).
agent=read16(root/'demo_de.wav',8)
user=read16(root/'humaneness-small/full-ft-speaker-study/audio/diagnostic-idris-natural_duration-0.mp3',4)
n=16000*8;agent=np.pad(agent[:n],(0,max(0,n-len(agent))))
rows=[]
for name in ['silence','normal_conversation','bot_echo','user_interrupts']:
for echo_db in ([-26,-40,-60,-90] if name in ('bot_echo','user_interrupts') else [None]):
for pairing in (['arrival_fixed_150ms','timestamps'] if name in ('bot_echo','user_interrupts') else ['timestamps']):
await asyncio.sleep(.1)
session=ears.open_session(ears.defaults)
if session is None:raise RuntimeError('Close the browser microphone session before synchronization checks')
mic=np.zeros(n,np.float32);reference=np.zeros(n,np.float32);mode='listen'
if name=='normal_conversation':mic[:len(user)]=user
if name in ('bot_echo','user_interrupts'):
mode='speaking';mic=agent*10**(echo_db/20);reference=agent.copy()
if name=='user_interrupts':
at=24000;length=min(len(user),n-at);mic[at:at+length]+=user[:length]
if pairing=='arrival_fixed_150ms':
# Recreate a 550 ms actual arrival/buffer path minus a 150 ms fixed guess.
ahead=6400;reference=np.r_[reference[ahead:],np.zeros(ahead,np.float32)]
events=[];tick=time.perf_counter();peaks=[]
try:
for i in range(0,n,5120):
a=mic[i:i+5120];b=reference[i:i+5120]
if len(a)<5120:a=np.pad(a,(0,5120-len(a)));b=np.pad(b,(0,5120-len(b)))
out=await ears.feed(session,a,b,mode)
peaks.extend(e.get('yield_prob',0.) for e in out if e['kind']=='FRAME')
events.extend((e['kind'],(i+5120)/16000) for e in out if e['kind']!='FRAME')
finally:ears.close(session)
rows.append({'scenario':name,'pairing':pairing,'echo_db':echo_db,'seconds':time.perf_counter()-tick,
'endpoint_count':sum(k=='ENDPOINT' for k,t in events),'barge_count':sum(k=='BARGE' for k,t in events),
'barge_times_s':[t for k,t in events if k=='BARGE'],
'max_yield':max(peaks,default=0.),
'contains_transcripts':False})
return {'conditions':rows,'note':'Running Nemotron/TYGER; synthetic fixtures, reconstructed arrival-order control, not a measured before-state on the lost machine. Hardware microphone accuracy requires physical testing.'}
|