voxi-lite / tests /latency.py
Joshuant's picture
Voxi Lite v0.1.1: low-latency clause streaming (speed-adaptive chunks), phrase cache warm(), tests/latency.py
568bf85 verified
Raw History Blame Contribute Delete
4.5 kB
"""Voxi Lite streaming latency on this machine: time to first audio and playback underruns, sentence vs clause
chunking, plus the phrase cache.
python tests/latency.py [--threads 4] [--voices payal,arjun] [--out latency.json]
For each reply, chunks are rendered in order (as stream() does) while a simulated player starts the moment the first
chunk is ready and plays chunks back to back. An underrun is a moment the player has finished and the next chunk is
not ready yet (the caller hears a gap); reported as count and total seconds.
"""
import argparse
import json
import sys
import time
from pathlib import Path
import numpy as np
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
from voxi import SR, VoxiLite # noqa: E402
REPLIES = [
("hi_bill", "hi", "नमस्ते, मैं Voxi से बोल रही हूँ। आपका बकाया ₹1,250 है, जो 15 तारीख़ तक भरना है। क्या मैं पेमेंट लिंक भेज दूँ?"),
("hi_order", "hi", "आपका ऑर्डर कल शाम तक पहुँच जाएगा और ट्रैकिंग लिंक आपको एसएमएस पर भेज दिया गया है ताकि आप उसे देख सकें।"),
("hi_short", "hi", "जी, बिल्कुल। मैं अभी चेक करती हूँ।"),
("hinglish", "hinglish", "Aapka refund process ho gaya hai, aur paise teen se paanch din mein aapke account mein aa jayenge."),
("ta_agent", "ta", "உங்கள் ஆர்டர் நாளை மாலைக்குள் வந்துவிடும், கண்காணிப்பு இணைப்பு SMS மூலம் அனுப்பப்பட்டுள்ளது. வேறு ஏதாவது உதவி வேண்டுமா?"),
("en_agent", "en", "Thank you for calling. Your refund of rupees 2,400 has been processed and it will reach your account in three to five working days."),
]
GREETING = "नमस्ते, मैं Voxi से बोल रही हूँ।"
def run(tts, text, lang, voice, chunking):
t0 = time.perf_counter(); ready, durs = [], []
for y in tts.stream(text, lang, voice, chunking=chunking):
ready.append(time.perf_counter() - t0); durs.append(len(y) / SR)
play_end, under, n_under = ready[0], 0.0, 0
for r, d in zip(ready, durs):
if r > play_end + 1e-3:
under += r - play_end; n_under += 1; play_end = r
play_end += d
return {"first_s": round(ready[0], 3), "chunks": len(ready), "audio_s": round(sum(durs), 2),
"underruns": n_under, "underrun_s": round(under, 3)}
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--threads", type=int, default=4); ap.add_argument("--voices", default="payal,arjun")
ap.add_argument("--out", default="")
a = ap.parse_args()
tts = VoxiLite(threads=a.threads)
tts.synthesize("नमस्ते।", "hi") # warm-up
rows = []
for voice in a.voices.split(","):
for name, lang, text in REPLIES:
for chunking in ("sentence", "clause"):
tts._cache.clear()
r = {"reply": name, "voice": voice, "chunking": chunking, **run(tts, text, lang, voice, chunking)}
rows.append(r); print(json.dumps(r, ensure_ascii=False), flush=True)
summ = {}
for chunking in ("sentence", "clause"):
rs = [r for r in rows if r["chunking"] == chunking]
f = [r["first_s"] for r in rs]
summ[chunking] = {"first_s_median": round(float(np.median(f)), 3), "first_s_p90": round(float(np.percentile(f, 90)), 3),
"first_s_max": round(max(f), 3), "replies_with_underrun": sum(r["underruns"] > 0 for r in rs),
"underrun_s_total": round(sum(r["underrun_s"] for r in rs), 2), "n": len(rs)}
tts._cache.clear()
tts.warm([GREETING], "hi", a.voices.split(","))
t0 = time.perf_counter(); next(tts.stream(GREETING, "hi", a.voices.split(",")[0])); cached = time.perf_counter() - t0
summ["cached_greeting_first_s"] = round(cached, 4)
print("SUMMARY", json.dumps(summ), flush=True)
if a.out:
Path(a.out).write_text(json.dumps({"summary": summ, "rows": rows}, indent=1, ensure_ascii=False), encoding="utf-8")
if __name__ == "__main__":
main()