Add PrimeTTS v2 (MB-iSTFT-VITS 34.7M, Xinran voice, 16 kHz) as default model

Single-ONNX end-to-end VITS served from Luigi/PrimeTTS v2_mbistft_16k/;
v1 pipeline kept for A/B. ORT-CPU parity vs torch verified (cos 1.0000).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X5N9x7k2c4iFcDzPKGcf2x

.gitignore

@@ -0,0 +1,3 @@
+__pycache__/
+*.pyc
+G2PWModel/

README.md

@@ -1,5 +1,5 @@
 ---
-title: Inflect-Nano-v1 vs PrimeTTS-8k — A/B TTS Demo
+title: PrimeTTS v2 — on-device zh-TW / English TTS
 emoji: 🗣️
 colorFrom: indigo
 colorTo: pink
@@ -7,7 +7,7 @@ sdk: docker
 app_port: 7860
 pinned: false
 license: apache-2.0
-short_description: 'A/B: original Inflect-Nano-v1 (EN) vs PrimeTTS-8k (zh-TW+EN)'
+short_description: 'PrimeTTS v2 (34.7M VITS, zh-TW+EN) vs v1 & baselines'
 models:
 - owensong/Inflect-Nano-v1
 - Luigi/PrimeTTS

app.py

@@ -31,7 +31,8 @@ INFLECT = "Inflect-Nano-v1 · 4.63M · FastSpeech+Snake-HiFiGAN — English, 24
 BREEZE2 = "Breeze2-VITS"
 MELO = "MeloTTS"
 MATCHA = "Matcha"
-V1B = "PrimeTTS"
+V1B = "PrimeTTS v1"
+V2 = "PrimeTTS v2"
 
 
 def host_regulate(cond, dur, pitch, abs_bins, max_frames):
@@ -109,6 +110,35 @@ def _v1b_engine():   # the flagship: 16 kHz acoustic (w/ frame_pitch_refiner) +
     return _ENG[V1B]
 
 
+def _v2_engine():   # PrimeTTS v2: MB-iSTFT-VITS 34.7M end-to-end (Xinran voice), 16 kHz, single ONNX
+    if V2 in _ENG: return _ENG[V2]
+    from huggingface_hub import hf_hub_download
+    import frontend_bopomofo as F
+    p = hf_hub_download(PRIMETTS_REPO, "v2_mbistft_16k/primetts_v2_xinran.onnx")
+    F.text_to_ids("您好")
+    _ENG[V2] = dict(kind="mbvits", F=F, sr=16000, sess=_ort(p))
+    return _ENG[V2]
+
+
+def _synth_v2(eng, text):
+    """Single-pass VITS: text ids (blank-interleaved, add_blank training) -> wav. Windowed
+    rel-pos attention has no absolute-position saturation, so no chunking is needed."""
+    o = eng["F"].text_to_ids(text)
+    if not o["phone_ids"]:
+        raise gr.Error("No pronounceable content found in the text.")
+    def blank(seq):
+        out = [0] * (2 * len(seq) + 1)
+        out[1::2] = seq
+        return np.array([out], np.int64)
+    x, tn, lg = blank(o["phone_ids"]), blank(o["tone_ids"]), blank(o["lang_ids"])
+    wav = eng["sess"].run(None, {
+        "x": x, "tone": tn, "lang": lg,
+        "x_lengths": np.array([x.shape[1]], np.int64),
+        "noise_scale": np.array([0.667], np.float32),
+        "length_scale": np.array([1.0], np.float32)})[0]
+    return wav.reshape(-1)
+
+
 def _sherpa_vits(repo, model_fn):
     import sherpa_onnx
     from huggingface_hub import snapshot_download
@@ -144,7 +174,7 @@ def _matcha_engine():
     return _ENG[MATCHA]
 
 
-ENGINES = {V1B: _v1b_engine}  # latest PrimeTTS only (encoder/decoder/vocoder pipeline)
+ENGINES = {V2: _v2_engine, V1B: _v1b_engine}  # PrimeTTS v2 (single-ONNX VITS) + v1 (enc/dec/voc pipeline)
 SHERPA = {BREEZE2: _breeze2_engine, MELO: _melo_engine, MATCHA: _matcha_engine}  # sherpa-onnx end-to-end
 
 
@@ -207,7 +237,8 @@ def generate(model: str, text: str) -> str:
             raise gr.Error("No audio produced — try Traditional-Chinese (and/or English) input for this model.")
     else:
         eng = ENGINES[model]()
-        wav = _synth_inflect(eng, text); sr = eng["sr"]
+        wav = _synth_v2(eng, text) if eng["kind"] == "mbvits" else _synth_inflect(eng, text)
+        sr = eng["sr"]
     peak = float(np.max(np.abs(wav)))      # peak-normalize -> consistent, loud output across models
     if peak > 1e-6:
         wav = wav * (0.97 / peak)
@@ -218,12 +249,14 @@ def generate(model: str, text: str) -> str:
 
 
 DESCRIPTION = """
-**PrimeTTS** is a ~4M-parameter zh-TW + English text-to-speech that runs entirely on **CPU** (built for the Jetson Nano).
-Pick a model, type text, hit **Generate** — and hear how a tiny on-device model compares to systems 6–8× larger.
+**PrimeTTS** is an on-device zh-TW + English text-to-speech family built for the Jetson Nano.
+**v2** (new) is a 34.7M end-to-end MB-iSTFT-VITS targeting the Nano's **GPU**; **v1** is the 5M
+FastSpeech pipeline for pure-CPU. Pick a model, type text, hit **Generate**.
 
 | Model | Params | Architecture | Rate | Languages |
 |:--|:--|:--|:--|:--|
-| **PrimeTTS** · *default* | **~5.0M** | FastSpeech + Snake-HiFiGAN (+ pitch refiner) | 16 kHz | zh-TW + English |
+| **PrimeTTS v2** · *default* | **34.7M** | MB-iSTFT-VITS (end-to-end) | 16 kHz | zh-TW + English |
+| PrimeTTS v1 | ~5.0M | FastSpeech + Snake-HiFiGAN (+ pitch refiner) | 16 kHz | zh-TW + English |
 | Breeze2-VITS | ~30M | VITS (end-to-end) | 22 kHz | zh-TW |
 | MeloTTS | ~32M | VITS | 8 kHz | zh + English |
 | Matcha | ~26M | Flow-matching + Vocos | 8 kHz | zh-TW + English |
@@ -231,8 +264,9 @@ Pick a model, type text, hit **Generate** — and hear how a tiny on-device mode
 
 ARTICLE = """
 ---
-**PrimeTTS is ~6–8× smaller** than the others — the point of this A/B is how close a tiny on-device CPU model gets to the
-big VITS / Matcha systems. Measured on a Jetson Nano (ORT-CPU): **CER 0.117**, **RTF 0.347**.
+**PrimeTTS v2** is a from-scratch MB-iSTFT-VITS (conv-only, no LSTM) distilled from a VibeVoice-Large
+teacher; on our 36-sentence zh/mix/en eval its X-ASR CER is **0.027** (below the teacher's 0.043).
+**v1** shows how far a ~5M pure-CPU model gets (Jetson Nano ORT-CPU: CER 0.117, RTF 0.347).
 
 <sub>First use of each model downloads its weights — give it a minute. ·
 [Luigi/PrimeTTS](https://huggingface.co/Luigi/PrimeTTS) ·
@@ -244,7 +278,7 @@ big VITS / Matcha systems. Measured on a Jetson Nano (ORT-CPU): **CER 0.117**, *
 demo = gr.Interface(
     fn=generate,
     inputs=[
-        gr.Dropdown(choices=[V1B, BREEZE2, MELO, MATCHA], value=V1B, label="Model"),
+        gr.Dropdown(choices=[V2, V1B, BREEZE2, MELO, MATCHA], value=V2, label="Model"),
         gr.Textbox(label="Text", value="Anderson 先生您好,您 2024年3月15日 訂的 3 件商品總共 NT$1,299,���號 AB1234CD,發票已寄到 anderson.wang@gmail.com;有問題請撥 0912-345-678 轉分機 2580,或到台北市信義區松高路11號5樓櫃台,今天氣溫 28°C,降雨機率 70%,謝謝。", lines=3),
     ],
     outputs=gr.Audio(label="Generated audio", type="filepath"),
@@ -254,15 +288,15 @@ demo = gr.Interface(
     theme=gr.themes.Soft(),
     examples=[
         # pure English
-        [V1B, "Thank you for calling. How may I help you today?"],
+        [V2, "Thank you for calling. How may I help you today?"],
         # pure zh-TW
-        [V1B, "您好,這裡是台灣電力公司客服中心,很高興為您服務。"],
+        [V2, "您好,這裡是台灣電力公司客服中心,很高興為您服務。"],
         # zh-TW + English code-mix
-        [V1B, "幫您轉接給 Jason 王經理,他的分機是 2580。"],
+        [V2, "幫您轉接給 Jason 王經理,他的分機是 2580。"],
         # complex + representative: entity-rich customer-service (date/price/serial/email/phone)
-        [V1B, "Anderson 先生您好,您 2024年3月15日 訂的 3 件商品總共 NT$1,299,序號 AB1234CD,發票已寄到 anderson.wang@gmail.com,有問題請撥 0912-345-678,謝謝。"],
+        [V2, "Anderson 先生您好,您 2024年3月15日 訂的 3 件商品總共 NT$1,299,序號 AB1234CD,發票已寄到 anderson.wang@gmail.com,有問題請撥 0912-345-678,謝謝。"],
         # complex + representative: code-mix with date/time/count/address
-        [V1B, "會議在 2024年3月15日下午三點,共 5 位 guest 出席,地點在台北市信義區松高路11號5樓。"],
+        [V2, "會議在 2024年3月15日下午三點,共 5 位 guest 出席,地點在台北市信義區松高路11號5樓。"],
     ],
     cache_examples=False,
     flagging_mode="never",