Add TeraTTSv2 Gradio demo

README.md

@@ -22,6 +22,9 @@ For installation, Python examples, ONNX options, and model downloads, visit
 The demo also displays the normalized text sent to the model after number
 expansion, automatic Russian stress marking, and Unicode normalization.
 
+`ru_f1` and `ru_m5` are the recommended Russian speaker prompts and appear
+first in the Space selector. Language tags are required for synthesis.
+
 > **Important:** `<ru>…</ru>` text receives automatic stress markers by
 > default. For an English voice speaking Russian, try a `duration_scale` below
 > `1`—`0.8` is a useful starting point.

app.py

@@ -4,6 +4,7 @@ from __future__ import annotations
 
 import os
 import logging
+import unicodedata
 from functools import lru_cache
 
 import numpy as np
@@ -51,7 +52,11 @@ def synthesize(
     except Exception as error:
         LOGGER.exception("TeraTTS synthesis failed")
         raise gr.Error(f"Synthesis failed: {error}") from error
-    return (44_100, waveform), normalized_text
+    # ONNX receives NFKD because that is how its vocabulary was trained.  Show
+    # users the identical text in normal composed Unicode so ``й``/``ё`` stay
+    # legible in the browser instead of appearing as a letter plus a detached
+    # combining mark.
+    return (44_100, waveform), unicodedata.normalize("NFC", normalized_text)
 
 
 demo = gr.Interface(
@@ -70,10 +75,10 @@ demo = gr.Interface(
         ),
         gr.Dropdown(
             [
-                ("Russian female 1", "ru_f1"),
+                ("★ Best: Russian female 1", "ru_f1"),
+                ("★ Best: Russian male 5", "ru_m5"),
                 ("Russian female 2", "ru_f2"),
                 ("Russian male 1", "ru_m1"),
-                ("Russian male 5", "ru_m5"),
                 ("English female 3", "eng_f3"),
                 ("English female 4 — whisper", "eng_f4_whisper"),
                 ("English female 5", "eng_f5"),
@@ -109,7 +114,7 @@ demo = gr.Interface(
     ],
     outputs=[
         gr.Audio(label="TeraTTSv2", type="numpy"),
-        gr.Textbox(label="Normalized text sent to model", lines=5, interactive=False),
+        gr.Textbox(label="Normalized text", lines=5, interactive=False),
     ],
     title="TeraTTSv2",
     description="""
@@ -117,14 +122,18 @@ demo = gr.Interface(
 
 - **Use TeraTTS locally:** visit [TeraSpace/TeraTTSv2](https://huggingface.co/TeraSpace/TeraTTSv2)
   for installation, Python examples, ONNX options, and downloadable model files.
+- **Recommended speakers:** **Russian female 1** and **Russian male 5** are
+  the preferred voice prompts and appear first in the selector.
 - **Russian stress is automatic:** text inside `<ru>…</ru>` receives stress
   markers by default; any manual `+` marker is preserved.
 - **English voice speaking Russian:** lower **Duration scale** below `1`
   (start around `0.8`) and adjust by ear.
 
-Use `<en>…</en>` or `<ru>…</ru>`. Distilled is faster; Teacher uses the
-higher-fidelity sampler. Dictionary-only stress saves memory by skipping
-RUAccent neural models. Numbers are spelled out from their language tags.
+Language tags are required: use `<en>…</en>` or `<ru>…</ru>`. Distilled is
+faster; Teacher uses the higher-fidelity sampler. Dictionary-only stress saves
+memory by skipping RUAccent neural models. Numbers are spelled out from their
+language tags. Unsupported characters are skipped and reported in the Space
+logs.
 """,
 )