Public preview with no SLA: MAI voices vs Sume's sonic-preview beta

Both ecosystems have a preview tier you should keep out of production. Microsoft's MAI voices page says the models are in public preview with no SLA and are not recommended for production; its streaming transcription page says the same. On Sume's TTS router, sonic-preview is described as a provider beta channel whose output and availability can change without notice, and sonic-latest is an alias for sonic-3.6 that is never sonic-preview.

What each page says about the tier

The Microsoft voices page, updated 2026-10-01, states public preview, no SLA, and not recommended for production. The realtime transcription page also lists the model as public preview with no SLA. Neither page promises that voices, styles or regions will stay as they are.

Sume's TTS router catalog lists sonic-3.6, sonic-3.5, sonic-3, sonic-latest and sonic-preview, and GET /v1/tts-router/models returns them. The catalog notes for sonic-preview add one more limit: it is not compatible with pro voice clones, and those requests fail with voice_model_mismatch.

How to choose

Use the preview for auditions and tests. For a series that must sound the same next month, pin a stable id such as sonic-3.6; the alias sonic-latest moves when the provider ships a new stable release, which is the right behavior for a general tool and the wrong one for a continuing narrator.

Limits: this compares the labels each vendor gives its tier, not quality. A preview model may well sound better than a stable one, and that is exactly when the temptation to ship it is strongest.

A rule you can encode

Put the decision in code, so nobody ships a preview model by accident. This guard refuses any model id containing preview when the environment is production, and prints the Sume alias that stays on the stable release.

Keep the id in configuration, not in the request code, and log it with each job so a narration series can always be traced to the model that made it.

import os

def pick_tts_model(requested):
    env = os.environ.get("APP_ENV", "dev")
    if env == "production" and "preview" in requested:
        raise SystemExit(f"{requested!r} is a preview channel; use sonic-latest")
    return requested

print(pick_tts_model("sonic-latest"))
print(pick_tts_model("sonic-preview"))

Pre-launch checklist for any preview model

If you decide to use a preview anyway, treat it as a dependency that can disappear.

• Render and archive the audio you publish, so a later change to the model cannot alter a finished video.
• Keep a fallback model id in configuration and test the switch once before launch.
• Record the model id with every job, so a drift in sound can be traced.
• Do not promise a customer a voice you cannot re-render on a stable model.