March 2026: TTS v3 prosody model, video generation quality pass, and memory retrieval improvements

TTS v3: context-aware prosody model

• New prosody model predicts pitch, rate, and energy from semantic context rather than a fixed style.
• Naturalness score (5-point MOS scale, internal panel of 50 raters) improved from 3.6 to 4.1.
• Handles punctuation-free input more gracefully; fewer robotic pauses on long responses.
• TTFA unchanged at 0.9s p50.

Video generation: motion coherence and frame interpolation

• Replaced single-pass generation with a two-stage model: keyframe generation followed by frame interpolation.
• Temporal consistency score improved from 0.79 to 0.86 on internal benchmark.
• Inter-frame flickering on hair and clothing reduced by 61% vs. v1.1.
• p95 generation time reduced to 14s at 720p despite the two-stage pipeline.

Image generation: session-level style memory

• Character appearance attributes (clothing, hairstyle, accessories) now persisted as structured metadata per session.
• Image generation model receives this metadata at inference time to maintain visual continuity.
• Reduces prompt re-specification needed to maintain a consistent look across multiple images.

Ref: PHO-88

Memory retrieval latency: p50 under 80ms

• Reduced memory retrieval p50 from 140ms to 74ms by switching to an HNSW approximate-nearest-neighbour index.
• Index now rebuilt incrementally on write rather than on a nightly batch job.
• No changes to the embedding model or vector dimensions.

Voice call 4G drop fix

• Added 12kbps and 8kbps rungs to the WebRTC bitrate ladder.
• Deployed March 12; call drop rate on 4G connections down 74% vs. February baseline.
• Validated across 18 carrier/device combinations in the device lab.