Publish v2 deployment variant / card update: README.md

README.md

@@ -47,7 +47,19 @@ Two model generations:
 | **Sample rate** | 16 kHz |
 | **Voice** | female Mandarin, "Xinran" |
 | **Eval (36 held-out zh/mix/en sentences)** | X-ASR CER **0.027** overall — zh 0.033 · code-mix 0.039 · en 0.008 (below its 7B teacher's 0.043 on the same eval) |
-| **Runtime** | single ONNX (`primetts_v2_xinran.onnx`, ORT-CPU, RTF ~0.01 on a desktop core) · `primetts_v2_xinran.gguf` for the ggml/CUDA Jetson-Nano runtime (port in progress) |
+| **Runtime** | single ONNX (`primetts_v2_xinran.onnx`, ORT-CPU) · `primetts_v2_xinran.gguf` for the ggml-CUDA Jetson-Nano runtime ([RapidSpeech.cpp](https://github.com/vieenrose/RapidSpeech.cpp), `mbistft-vits` arch) |
+
+### On-device deployment (measured on a Jetson Nano gen-1, Tegra X1)
+
+Two runtime tiers, both real-time. RTF = compute-time / audio-time (lower is faster; <1.0 = real-time).
+
+| Tier | Runtime | Precision | RTF | Quality |
+|---|---|---|---|---|
+| **GPU** | RapidSpeech.cpp ggml-CUDA, 1 CPU thread | fp32 / fp16 | **0.42** (2.4× RT) | full (parity 0.9998) |
+| **CPU** | onnxruntime, 4 threads | **fp32** | **0.52** (1.9× RT) | full — voice-cos 0.916, CER 0.033 |
+| **CPU** | onnxruntime, 4 threads | **q8** (int8) | **0.27** (3.7× RT) | fast/small, but reduced voice fidelity (voice-cos 0.748, CER 0.048) — a **QAT int8** build to restore identity is in progress |
+
+Notes: the GPU RTF is launch-overhead-bound on Maxwell (sm_53, no CUDA-graph replay) — 0.42 is the practical floor there; the **GPU is not required** — CPU-only fp32 clears real-time with full fidelity. `q8` beats the GPU on speed but shifts timbre (int8); `fp16` ONNX is half-size but ORT-CPU casts it to fp32 so it gives **no CPU speedup** (size only). Pick fp32 for fidelity, q8 for smallest/fastest.
 
 **Training:** distilled from a **VibeVoice-Large** (MIT) teacher speaking the `zh-Xinran_woman` preset —
 29k utterances over the same entity-rich zh-TW corpus as v1, per-utterance speaker-consistency QC
@@ -147,8 +159,10 @@ and a small brand lexicon. Text past `max_frames` is auto-chunked at punctuation
 ## Model files
 
 ```
-v2_mbistft_16k/primetts_v2_xinran.onnx   ← PrimeTTS v2 FLAGSHIP (34.7M, 16 kHz, single ONNX) — the demo serves this
-v2_mbistft_16k/primetts_v2_xinran.gguf   ← same weights for the ggml/CUDA Jetson-Nano runtime
+v2_mbistft_16k/primetts_v2_xinran.onnx        ← PrimeTTS v2 FLAGSHIP fp32 (34.7M, 16 kHz) — full quality, the demo serves this
+v2_mbistft_16k/primetts_v2_xinran.gguf        ← fp32 weights for the ggml-CUDA Jetson-Nano runtime (RapidSpeech.cpp)
+v2_mbistft_16k/primetts_v2_xinran_fp16.onnx   ← fp16 ONNX (55 MB, half size, lossless quality; ORT-CPU no speedup — size only)
+v2_mbistft_16k/primetts_v2_xinran_q8.onnx     ← static-int8 ONNX (28 MB, fastest CPU RTF 0.27; reduced voice fidelity — QAT build coming)
 v1b_16k/{acoustic_encoder,acoustic_decoder,vocoder}.onnx + meta.json  ← v1 16 kHz (~5.0M, CPU)
 v1b_8k/ {acoustic_encoder,acoustic_decoder,vocoder}.onnx + meta.json  ← v1 leanest on-device (4.09M, 8 kHz, Nano CPU)
 {acoustic_encoder,…}.onnx + meta.json   ·   v3_4.6M/                  ← legacy 24 kHz variants (6.85M / 4.63M), for record