r/hold-slayer
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(lab): make the speech fixture actually classify as speech

Browse Source 
The speech fixture classified as LIVE_HUMAN on its first 3s window and drifted
to MUSIC for every window after. I had validated only the first window and
reported the fixture as verified, which overstated it: any lab result resting
on that fixture — the hold-slayer scenarios above all — was proving less than
it appeared to.

The cause was one modelling error, not a tuning problem. `_detect_tonality`
looks for an autocorrelation peak above 0.5 in the 50-1000 Hz lag range, and
each syllable used a *constant* f0, which is perfectly periodic there. That
scored is_tonal=True, handing the music score a free 0.3 that speech could not
outrun — and the decision requires speech_score to strictly exceed
music_score, so ties went to music.

Real voices glide and jitter, so the periodicity never locks. The fundamental
now follows a per-syllable pitch contour (rise or fall, plus ~2% cycle-to-cycle
jitter), with the frequency integrated to phase rather than multiplied by t —
`2*pi*f*t` is only a chirp when f is the instantaneous rate, which it is not
once f0 itself moves. is_tonal is now False in every window.

Two smaller fixes fell out of that:

- Aspiration noise is high-passed rather than broadband. Flat noise puts
  energy in every Goertzel bin, so the strongest DTMF row and column both
  clear the detector's `total_power * 0.1` threshold and each syllable reads
  as a keypress. A first-difference filter leaves the 697-1633 Hz bands
  comparatively empty. The level is set for margin — spectral flatness lands
  at ~0.46, mid-way through the 0.1-0.5 band, not on an edge.

- The music fixture gained two more harmonics and a recording-style noise
  floor. Windows straddling a chord change had a momentarily sparse spectrum
  and fell *below* the music score's 0.05 flatness floor, scoring as speech.

All three fixtures now classify correctly in 100% of windows (music 27/27,
speech 5/5, silence 2/2), and remain correct when the window is stepped by
half a window — a fixture that only works on aligned boundaries would still
be a trap in a live call, where the analysis window has no relationship to
where the audio began.

Confirmed on a real call through the lab: scenario 1003 now shows the whole
hold-slayer arc, speech -> sustained music -> speech, matching the dialplan.

tests/test_lab_fixtures.py guards this: it sweeps every window rather than
sampling the first, which is exactly what the original validation missed, and
checks the generator is byte-for-byte deterministic. It skips when the
fixtures have not been generated, since they are gitignored.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Robert Helewka
2026-07-29 07:56:58 -04:00
parent 2a05be27bf
commit 92c45e9c4d
2 changed files with 158 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

61
tests/lab/sounds/generate.py
View File

@@ -12,7 +12,6 @@ finds `lab-music.sln`.
python generate.py [outdir]
"""
import struct
import sys
from pathlib import Path
@@ -29,13 +28,14 @@ def _write_sln(path: Path, samples: np.ndarray) -> None:
print(f" {path.name}: {len(pcm) / RATE:.1f}s ({path.stat().st_size} bytes)")
def make_music(seconds: float = 30.0) -> np.ndarray:
def make_music(seconds: float = 30.0, seed: int = 7) -> np.ndarray:
"""Sustained multi-harmonic tones — what the classifier must call MUSIC.
A chord progression with stable pitch and strong harmonic structure. The
steady spectrum across a long window is what distinguishes music from
speech; this deliberately has no pauses.
"""
rng = np.random.default_rng(seed)
t = np.linspace(0, seconds, int(RATE * seconds), endpoint=False)
# A-minor-ish progression, one chord per 2s bar.
chords = [(220.0, 261.6, 329.6), (196.0, 246.9, 293.7),
@@ -48,12 +48,21 @@ def make_music(seconds: float = 30.0) -> np.ndarray:
break
mask = (t >= start) & (t < end)
for j, freq in enumerate(chord):
# Fundamental plus two harmonics, decaying — a plucked-string feel.
for h, amp in ((1, 0.30), (2, 0.12), (3, 0.05)):
# Fundamental plus four harmonics, decaying — a plucked-string
# feel. Enough harmonics to keep spectral flatness inside the
# music score's 0.05-0.4 band: with only three, some windows fall
# *below* 0.05 (too pure to read as music) and score as speech.
for h, amp in ((1, 0.30), (2, 0.12), (3, 0.05), (4, 0.03), (5, 0.02)):
out[mask] += amp / (j + 1) * np.sin(2 * np.pi * freq * h * t[mask])
# Gentle per-bar envelope so bars are distinguishable but never silent.
env = 0.8 + 0.2 * np.sin(2 * np.pi * (t[mask] - start) / bar)
out[mask] *= env
# Recording-style noise floor. Windows straddling a chord change have a
# momentarily sparse spectrum and land just *under* the music score's
# 0.05 flatness floor, scoring as speech. This is well below the level
# that would disturb tonality — every real recording has one.
out += rng.normal(0, 0.004, len(out))
return out * 0.45
@@ -74,16 +83,44 @@ def make_speech(seconds: float = 8.0, seed: int = 1337) -> np.ndarray:
mask = (t >= pos) & (t < pos + syl)
if mask.any():
local = t[mask] - pos
f0 = rng.uniform(95, 165) # fundamental — adult speaking range
# Two formants, swept slightly across the syllable. The ranges
# deliberately avoid the DTMF bands (rows 697-941, columns
# 1209-1633): a formant pair landing on both trips the Goertzel
# detector and the whole utterance is classified as a keypress.
f1 = rng.uniform(300, 620) + rng.uniform(-40, 40) * local / syl
f2 = rng.uniform(1750, 2600) + rng.uniform(-120, 120) * local / syl
sig = (0.50 * np.sin(2 * np.pi * f0 * local)
frac = local / syl
# Pitch CONTOUR, not a constant. This is the single feature that
# separates this fixture from music. `_detect_tonality` looks for
# an autocorrelation peak > 0.5 in the 50-1000 Hz lag range; a
# fixed f0 is perfectly periodic there, scores is_tonal=True, and
# hands the music score a free 0.3 that speech cannot outrun.
# Real voices glide and jitter, so the periodicity never locks.
f0_start = rng.uniform(95, 165)
f0_end = f0_start * rng.uniform(0.72, 1.38) # rise or fall
f0 = f0_start + (f0_end - f0_start) * frac
# Cycle-to-cycle jitter on top of the glide (~2% is human).
f0 *= 1.0 + 0.02 * rng.standard_normal(len(local))
# Integrate frequency to phase — with a varying f0, `2*pi*f*t`
# would be wrong (that is a chirp only if f is the *instantaneous*
# rate, which it is not once f0 itself moves).
ph0 = 2 * np.pi * np.cumsum(f0) / RATE
# Two formants, swept across the syllable. The ranges deliberately
# avoid the DTMF bands (rows 697-941, columns 1209-1633): a formant
# pair landing on both trips the Goertzel detector and the whole
# utterance is classified as a keypress.
f1 = rng.uniform(300, 620) + rng.uniform(-40, 40) * frac
f2 = rng.uniform(1750, 2600) + rng.uniform(-120, 120) * frac
sig = (0.50 * np.sin(ph0)
+ 0.30 * np.sin(2 * np.pi * f1 * local)
+ 0.18 * np.sin(2 * np.pi * f2 * local))
# Aspiration noise — HIGH-PASSED, not broadband. Real speech noise
# sits above the formants; flat noise puts energy in every
# Goertzel bin, so the strongest DTMF row and column both clear
# the detector's `total_power * 0.1` threshold and every syllable
# reads as a keypress. A first-difference filter (y[n]-y[n-1]) is
# a cheap +6dB/octave tilt that leaves the 697-1633 Hz DTMF bands
# comparatively empty. The 0.09 level is chosen for margin: it puts
# spectral flatness at ~0.46, mid-way through the 0.1-0.5 band the
# speech score rewards, rather than on either edge.
noise = rng.standard_normal(len(local) + 1)
sig += 0.09 * np.diff(noise)
# Raised-cosine envelope: no clicks at syllable edges.
sig *= np.sin(np.pi * local / syl) ** 0.6
out[mask] += sig