Voice API quotas (TTS / STT)

Voice providers charge by “number of seconds / characters per month”. Unlike LLMs, limits here are strict — exceeding = 403, and the user doesn’t receive audio. We count ourselves and alert in advance:

import os, requests, json, time

S = "/tmp/voice-quota.json"
LIMITS = {                                # characters / seconds per month
    "elevenlabs": 100_000,
    "deepgram":   45_000,
    "whisper":    60_000,
}

def observe(provider: str, units: int):
    s = (json.load(open(S)) if os.path.exists(S) else {})
    month = time.strftime("%Y-%m")
    s.setdefault(month, {}).setdefault(provider, 0)
    s[month][provider] += units
    json.dump(s, open(S, "w"))

    used  = s[month][provider]
    limit = LIMITS.get(provider, 1)
    for thr in (0.7, 0.9, 0.99):
        flag = f"/tmp/voice-{provider}-{month}-{int(thr*100)}.flag"
        if used / limit >= thr and not os.path.exists(flag):
            push(f"🎙️ {provider}: {int(thr*100)}% квоты",
                 f"{used:,} / {limit:,}",
                 9 if thr >= 0.9 else 5)
            open(flag, "w").close()

def push(t, m, p):
    requests.post(f"{os.environ['NOTIFLY_URL']}/message",
                  params={"token": os.environ["NOTIFLY_TOKEN"]},
                  json={"title": t, "message": m, "priority": p}, timeout=5)

At the same time — an HTTP monitor on api.elevenlabs.io/v1/user to detect quota_exceeded in their own response.