Files
Ryan Vogel 23baa80aaa Measure on-device Android dictation across four languages
Replace the phase-one single-phrase evidence with a fixed 40-clip multilingual, partly music-mixed corpus. Record per-clip errors, load, warm inference, memory, two-core sensitivity and delivered platform-service results separately. Keep an independent foreground editor and media track for mic-to-insertion verification.

Verified on an API 35 ARM64 Studio emulator: Whisper tiny/base/small, Parakeet TDT v3 int8 and strictly on-device Android recognition. Small in its short window made 15/285 word errors; the platform service returned 33/40. Both native runners also decoded the notebook clip after tightening the PCM input checks.
2026-11-20 12:26:12 -05:00

38 lines
1.9 KiB
Python

"""Generate the fixed 16 kHz/mono corpus on macOS. Music is synthesized, not sourced."""
import array, csv, hashlib, math, subprocess, sys, tempfile, wave
from pathlib import Path
dest = Path(sys.argv[1])
dest.mkdir(parents=True, exist_ok=True)
rate = 16000
with open(Path(__file__).with_name("corpus.tsv"), encoding="utf8") as src:
rows = list(csv.DictReader(src, delimiter="\t"))
for row in rows:
with tempfile.TemporaryDirectory(dir=dest) as tmp:
aiff = Path(tmp) / "voice.aiff"
subprocess.run(["say", "-v", row["voice"], "-r", row["wpm"], "-o", aiff, row["reference"]], check=True)
raw = subprocess.check_output(["ffmpeg", "-loglevel", "error", "-i", aiff, "-f", "s16le", "-ac", "1", "-ar", str(rate), "-"])
samples = array.array("h", raw)
if not samples or max(samples) < 500:
raise ValueError("Empty synthesis: " + row["id"])
voice_rms = math.sqrt(sum(s*s for s in samples)/len(samples))
samples = [0]*3200 + list(samples) + [0]*6400
if row["snr_db"] != "clean":
tones = (261.63, 329.63, 392.00, 493.88, 220.00, 261.63, 329.63, 440.00)
bed = []
for i in range(len(samples)):
t = i/rate
beat = int(t/0.24)
phase = t%0.24
f = tones[beat%len(tones)]
env = min(1.,phase/.012)*math.exp(-phase*8)
bed.append(env*(math.sin(2*math.pi*f*t)+.25*math.sin(4*math.pi*f*t))+.25*math.sin(2*math.pi*110*t))
gain = voice_rms / (10**(float(row["snr_db"])/20)*math.sqrt(sum(s*s for s in bed)/len(bed)))
samples = [a+b*gain for a,b in zip(samples,bed)]
gain = min(1.,30000/max(abs(s) for s in samples))
wav = dest / (row["id"]+".wav")
with wave.open(str(wav), "wb") as out:
out.setparams((1,2,rate,0,"NONE","not compressed"))
out.writeframes(array.array("h",(round(s*gain) for s in samples)).tobytes())
print(row["id"], len(samples), hashlib.sha256(wav.read_bytes()).hexdigest(), sep="\t")