docs(IndexTTS): describe reference conditioning

README.md

@@ -40,6 +40,9 @@
 - **2026-08** **Hybrid CPU/GPU expert placement reaches four architecture families.** Qwen3-MoE,
   Qwen3.5/3.6, Nemotron-H, and DeepSeek-V2 can run routed experts on the CPU while the rest of the
   model stays on the selected accelerator. The end-to-end token and speed gates are still pending.
+- **2026-08** **IndexTTS 2.5 now uses its reference audio.** The clip conditions both the talker
+  and S2Mel stages. Repeated runs with one clip are bit-identical, while different clips change
+  the output. Voice-cloning parity with vLLM-Omni remains unmeasured.
 - **2026-08** **MiniMax-Music3 generates music through the public API.** Every pipeline stage is
   implemented and gated. The server exposes it through `POST /v1/audio/speech`; no reference speed
   number is available yet.

docs/models/indextts-2-5.md

@@ -56,16 +56,21 @@ curl http://localhost:8000/v1/audio/speech \
 
 ## Current limitations
 
-The engine requires a reference clip but does not use the reference clip for
-conditioning. The clip does not select the output voice. Reference-audio
-conditioning remains incomplete.
+The reference clip conditions both halves of the model. The engine projects a
+CAMPPlus speaker vector into the talker's first conditioning row and passes the
+same vector to the S2Mel front end as its style input. The clip must use a 16
+kHz sample rate because the engine has no resampler.
+
+This wiring does not establish voice-cloning parity. Repeated runs with one
+clip are bit-identical, and different clips change the output, but vLLM-Omni is
+not pinned as an oracle. The inferred emotion path is also not implemented.
 
 The vLLM-Omni quality and parity comparison is pending. Current structural
-gates show that the pipeline renders audio. They do not establish output
-quality or parity with the oracle.
+gates show that the pipeline renders audio and that the reference path affects
+the result. They do not establish output quality or parity with the oracle.
 
-The inferred emotion path is not implemented. The request can provide text,
-but the server exposes no named voice or speaking-rate control.
+The request can provide text, but the server exposes no named voice or
+speaking-rate control.
 
 See the [server reference](../reference/server.md) for the shared endpoint
 contract. See the

docs/reference/server.md

@@ -69,9 +69,10 @@ returns a 44.1 kHz stereo WAV. See the
 
 IndexTTS-2.5 loads from `--speech-model` and returns a 22.05 kHz mono WAV. Each
 request must include `reference_audio` as a data URL for a 16-bit PCM mono WAV.
-The current engine does not use the reference clip for conditioning, so the
-clip does not select the output voice. The vLLM-Omni quality comparison is also
-pending. See the
+The clip must have a 16 kHz sample rate. Its CAMPPlus speaker vector conditions
+the talker and S2Mel stages. Different clips change the output, but this does
+not establish voice-cloning parity. The inferred emotion path is not
+implemented, and the vLLM-Omni quality comparison is pending. See the
 [owning model specification](../../.agents/specs/indextts-2-5.md) for the owned
 limitations and verification evidence.