Audra Pro (MOSS): GPU narration when Essential feels flat — Audra

Essential presets are clean and consistent, and that is exactly the problem for some scripts: ads, character reads, and premium audiobook hosting need performance, not just clarity. audra-pro runs GPU MOSS inference with instruct control at a 2.0x multiplier ($50/1M prepaid) — the engine you reach for when the voice has to sell something or act.

This post is the Pro decision: what MOSS gives you, the instruct-control model, what it costs at real ad/audiobook scale, and when it is wasted money.

What Pro changes

Same API, model audra-pro. GPU MOSS inference produces more expressive delivery — emphasis, energy variation, character voicing — and accepts instruction-style control to shape the read. Delivery profiles and podcast pacing apply as on other engines. The 50k sync / 500k jobs limits and QA-gated billing are unchanged.

curl
curl https://audratalks.com/v2/speech \
  -H "Authorization: Bearer sk-audra-..." \
  -H "Content-Type: application/json" \
  -d '{"model":"audra-pro","voice":"<pro-voice-slug>","text":"Sixty seconds. One decision. Switch today.","format":"mp3"}' \
  --output ad-read.mp3
Python
import requests

resp = requests.post(
    "https://audratalks.com/v2/speech",
    headers={"Authorization": "Bearer sk-audra-..."},
    json={"model": "audra-pro", "voice": pro_slug, "text": ad_script, "format": "mp3"},
    timeout=3600,
)
resp.raise_for_status()
JavaScript
const res = await fetch("https://audratalks.com/v2/speech", {
  method: "POST",
  headers: { Authorization: "Bearer sk-audra-...", "Content-Type": "application/json" },
  body: JSON.stringify({ model: "audra-pro", voice: proSlug, text: adScript, format: "mp3" }),
});
if (!res.ok) throw new Error(await res.text());

Where Pro earns 2x

Where Pro is wasted

Long-form factual narration at scale: a 250k-character audiobook costs $12.50 on Pro versus $6.25 on Essential, and for nonfiction reference material the quality difference is not audible enough to double the budget. The honest heuristic: Pro for scripts measured in hundreds of characters where performance is the product; Essential for scripts measured in thousands where information is.

Failure mode: quality regression by assumption

Teams sometimes flip the whole pipeline to Pro expecting across-the-board improvement and hear none — or hear delivery that no longer matches episodes rendered on Essential. Voice catalog and timbre differ per engine, so a wholesale model switch mid-series changes the show's sound. The migration pattern: shadow-render a representative script set on Pro, A/B them blind, and only promote the segments that measurably win.

Cost math at ad scale

A weekly show selling four ad slots: 4 x 600 chars x 52 weeks = 125k characters/year = $6.25/year on Pro. Ad inventory is so cheap per-character that engine choice is a creative decision, not a budget one. Model larger volumes on /calc.

The full engine ladder: the four engines guide. Multi-speaker drama (a different tool than Pro): the Conversations post.