docs: voice cloning is supported — Mimi encoders are included (fixes FluidAudio#877) (#4)

feat(voice-cloning): per-language mimi_encoderv3 for spanish_24l (#793)

feat(voice-cloning): per-language mimi_encoderv3 for spanish_24l (#793)

feat(voice-cloning): per-language mimi_encoderv3 for portuguese_24l (#793)

feat(voice-cloning): per-language mimi_encoderv3 for portuguese_24l (#793)

feat(voice-cloning): per-language mimi_encoderv3 for italian_24l (#793)

feat(voice-cloning): per-language mimi_encoderv3 for italian_24l (#793)

feat(voice-cloning): per-language mimi_encoderv3 for german_24l (#793)

feat(voice-cloning): per-language mimi_encoderv3 for german_24l (#793)

feat(voice-cloning): per-language mimi_encoderv3 for french_24l (#793)

feat(voice-cloning): per-language mimi_encoderv3 for french_24l (#793)

feat(voice-cloning): per-language mimi_encoderv3 for spanish (#793)

feat(voice-cloning): per-language mimi_encoderv3 for spanish (#793)

feat(voice-cloning): per-language mimi_encoderv3 for portuguese (#793)

feat(voice-cloning): per-language mimi_encoderv3 for portuguese (#793)

feat(voice-cloning): per-language mimi_encoderv3 for italian (#793)

feat(voice-cloning): per-language mimi_encoderv3 for italian (#793)

feat(voice-cloning): per-language mimi_encoderv3 for german (#793)

feat(voice-cloning): per-language mimi_encoderv3 for german (#793)

Add speaker_proj_weight (24L) + encoder_recover_pinv for voice-clone reprojection (FluidAudio #793)

docs: PR#679 listening samples (same text+seed across gpu/ane/ane-state placements)

feat(v2.1/english): pocket_state multifunction MLState package (write_state/prefill/generate over shared KV state; mobius Trial 23). Loaded by FluidAudio PocketTtsModelPlacement.aneState (macOS 15+/iOS 18+); Swift-measured -14.9% e2e vs .ane, WER 0.

feat(v2.1): ANE-eligible flowlm_step_ane + cond_prefill_ane for all 9 remaining language packs (german/italian/portuguese/spanish 6L + french/german/italian/portuguese/spanish 24L). Rank-4 split-KV, scatter-free, explicit output names, no NaN-BOS protocol (mobius Trials 19/20). flowlm 100% ANE on every pack incl. 24L (1830 ops). Loaded by FluidAudio PocketTtsModelPlacement.ane.

feat(v2.1/english): ANE-eligible flowlm_step_ane + cond_prefill_ane (rank-4 split-KV, scatter-free; mobius Trials 19/20). flowlm 100% ANE / cond_prefill 92% under cpuAndNeuralEngine; bit-identical math to the rank-5 graphs; explicit output names; no NaN-BOS protocol (host passes the BOS latent directly). Loaded by FluidAudio PocketTtsModelPlacement.ane.

Add v2.1 english (optimized: fused decoder ANE + cond prefill + fp16 flowlm; same weights as v2) (#3)

Delete v2/spanish_24l/verify.wav

Delete v2/spanish/verify.wav

Delete v2/portuguese_24l/verify.wav

Delete v2/portuguese_24l/vera.safetensors

Delete v2/portuguese_24l/stuart_bell.safetensors

Delete v2/portuguese_24l/rafael.safetensors

Delete v2/portuguese_24l/peter_yearsley.safetensors

Delete v2/portuguese_24l/paul.safetensors

Delete v2/portuguese_24l/michael.safetensors

Delete v2/portuguese_24l/mary.safetensors

Delete v2/portuguese_24l/marius.safetensors

Delete v2/portuguese_24l/lola.safetensors

Delete v2/portuguese_24l/juergen.safetensors

Delete v2/portuguese_24l/jean.safetensors

Delete v2/portuguese_24l/javert.safetensors

Delete v2/portuguese_24l/jane.safetensors

Delete v2/portuguese_24l/giovanni.safetensors

Delete v2/portuguese_24l/george.safetensors

Delete v2/portuguese_24l/fantine.safetensors

Delete v2/portuguese_24l/eve.safetensors

Delete v2/portuguese_24l/estelle.safetensors

Delete v2/portuguese_24l/eponine.safetensors

Delete v2/portuguese_24l/cosette.safetensors

Delete v2/portuguese_24l/charles.safetensors

Delete v2/portuguese_24l/caro_davy.safetensors