Microsoft's new deepfake tool lets avatars speak for you

Microsoft just turned deepfake creation into an enterprise cloud feature.

At its Ignite 2023 conference, the tech giant introduced an Azure AI Speech tool called text-to-speech avatar, now available in public preview. The service lets users upload a photo of a person, feed it a script, and generate a photorealistic talking video of that person saying things they never actually uttered.

One minute of audio to clone your voice

The setup involves two synchronized models. One model animates the visual likeness to match the speech, while another synthetic voice model reads the script aloud.

These avatars speak multiple languages and can even hook into OpenAI's GPT-3.5 to field off-script customer questions in conversational chatbot setups. Microsoft pitches the tool for corporate training videos, product explainers, and customer service reps.

Alongside the avatars, Microsoft also debuted "personal voice" within its custom neural voice platform. It requires just a single one-minute audio sample to replicate someone's voice in seconds. It is designed for multi-language dubbing, custom audiobooks, and bespoke narrations.

Red tape and guardrails

Given how tools from rivals like Synthesia have been abused for political propaganda, Microsoft is wrapping the release in strict bureaucratic barriers.

Most Azure users only get access to generic, prebuilt avatars. Creating a custom avatar requires going through a limited-access registration process. Customers must prove they obtained explicit written consent from the person being modeled, with clear terms on duration and usage, and are required to label all resulting videos as AI-generated.

For voice cloning, Microsoft bans prerecorded audio. Users must record a specific consent statement live, which Microsoft verifies against the training sample. Synthetic voices also get watermarked, though detecting those watermarks requires applying for Microsoft's proprietary detection service.

Why it matters

We think this is a classic high-reward, high-anxiety launch. Dropping photorealistic cloning tech right as Hollywood just wrapped up a historic strike over digital likeness rights shows how aggressive big tech is willing to be.

For businesses, the cost savings are undeniable. Producing localized training modules or personalized support avatars usually drains thousands of dollars in studio time. Cutting that workflow down to a text prompt and an image upload is huge.

However, gating the technology behind corporate forms feels like a flimsy long-term shield. Once the barrier to convincing impersonation drops to a one-minute voice clip and a snapshot, enforcing terms of service across thousands of downstream clients becomes nearly impossible. By pushing synthetic humans into everyday business software, Microsoft is accelerating an internet where nobody can trust their own eyes and ears.