A dubbing engine built for performance, not just speech.

An explicit, controllable emotion bank.

Thirteen named emotions with a continuous intensity dial (mild to extreme), settable per line or inline mid-sentence. Sustained emotions can be layered, and source-audio emotion can be auto-detected and preserved through translation instead of flattened. Every setting is reproducible: dial it in once, regenerate the same performance every time.

Inline syntax: "[happy] So good to see you! [sad:strong] I just wish it were better news." Intensity levels are mild · moderate · strong · extreme, or any value 0–2.

Accent control with adjustable intensity.

Blend a speaker’s delivery toward a target accent with a 0–100% intensity slider while their voice identity is preserved. The blend operates on the speech-delivery conditioning only, never the timbre, so the speaker still sounds like themselves.

Accent centroids ship per language after passing a speaker-similarity guardrail (≥ 0.60 at half strength).

Zero-shot cloning, hardened for real references.

No per-voice training. References are assembled automatically from a speaker’s best segments, denoised adaptively, and quality-scored (duration, SNR, clipping) with warnings surfaced before you dub, so even a speaker with seconds of usable audio gets the best possible clone.

Timing that survives the edit.

Slot-aware synthesis picks a speaking rate so every line fills its original screen time, a decode-time rate controller steers token budgets, and DubScore grades the finished dub segment by segment. Weak lines are flagged and regenerable individually, without reprocessing the timeline.

Live dubbing with an honest latency meter.

A streaming ASR → translation → cloned-TTS cascade driven by one latency knob (conversational, balanced, broadcast). Session voices are pre-warmed and cached on the GPU, chunks synthesize over a priority lane, and the console shows the measured end-to-end lag next to the target, not an aspirational number.

Consent, watermarking and provenance built in.

Every clone requires a consent reference: dual-signed marketplace agreements, per-project actor consent, or an attested release. Output audio carries an inaudible watermark, C2PA manifests are signed, and a consent-vault report collates the full evidence chain for a studio’s compliance team.