Let the TTS server be told what language to speak

Every generation branch hardcoded language="english", and the endpoint
accepted no language parameter at all. German, French and the rest were
synthesised with English phonology and prosody - which is why the German
samples sounded like an English narrator reading German.

- /v1/tts now accepts a language form field, passed to all four branches.
  Default stays "english" so the existing English pipeline is unchanged;
  the renderer sends the target language from the model's own
  codec_language_id keys (exactly ten: chinese english french german
  italian japanese korean portuguese russian spanish).
- A/B on one German line with the same cloned voice: language=english
  gives 4.56s, language=german 4.08s, waveforms differing by 151% of
  signal amplitude. Audibly different renditions, not a no-op.

Also documented the ICL memory finding in the server comments: the branch
that consumes a cached prompt leaves x_vector_only_mode at its default of
False, so it decodes the 8k ref_code matrix on every request. One such
request takes the footprint from 5.4GB to 19.1GB and does not return. The
ref_audio branch passes x_vector_only_mode=True and is stable - 120 clips,
0 failures, 13.1GB.

Note the language fix does NOT address that; it was tested and still
blows up with language=german. Separate problem.

qwen_tts_server.py stays at its 8 pre-existing ruff errors. Suite: 721
passed, 8 pre-existing.

scripts/qwen_tts_server.py

@@ -403,6 +403,12 @@ async def set_attn(request: AttnRequest):
 async def generate_tts(
     text: str = Form(...),
     sample_rate: int = Form(24000),
+    # Was hardcoded to "english" in every branch, so German and French text was
+    # synthesised with English phonology and prosody - an English-accented
+    # foreign voice. Defaults to "english" so the existing English pipeline is
+    # byte-for-byte unchanged; multilingual callers pass their own language.
+    # "Auto" is also accepted and lets the model detect it.
+    language: str = Form("english"),
     voice_description: str | None = Form(None),
     ref_audio: UploadFile | None = File(None),
     ref_text: str | None = Form(None),
@@ -455,7 +461,7 @@ async def generate_tts(
                 wavs, sr = model.generate_voice_clone(
                     text=text,
                     voice_clone_prompt=items,
-                    language="english",
+                    language=language,
                     **kw,
                 )
                 return wavs, sr
@@ -480,7 +486,7 @@ async def generate_tts(
                         "x_vector_only_mode": [True],
                         "icl_mode": [False],
                     },
-                    language="english",
+                    language=language,
                     **gen_kwargs,
                 )
                 return wavs, sr
@@ -512,7 +518,7 @@ async def generate_tts(
                     ref_audio=(ref_audio_array, sample_rate),
                     ref_text=None,
                     x_vector_only_mode=True,
-                    language="english",
+                    language=language,
                     **kw,
                 )
                 return wavs, sr
@@ -524,7 +530,7 @@ async def generate_tts(
                 wavs, sr = model.generate_voice_design(
                     text=text,
                     instruct=voice_description,
-                    language="english",
+                    language=language,
                     **gen_kwargs,
                 )
                 return wavs, sr

scripts/render_lang_test.py

@@ -47,6 +47,13 @@ from umbraaudio.config import load_config  # noqa: E402
 
 TTS_URL = "http://127.0.0.1:8881"
 
+# Qwen3-TTS language names, from the model's own codec_language_id keys.
+LANG_CODES = {
+    "EN": "english", "ZH": "chinese", "JA": "japanese", "KO": "korean",
+    "DE": "german", "ES": "spanish", "FR": "french", "RU": "russian",
+    "IT": "italian", "PT": "portuguese",
+}
+
 # Chapter 1 of Deal Hunter is the voice-introduction sample.
 CHAPTER_ID = 1
 TRANSLATION_DIRNAME = "Lang_test"
@@ -103,6 +110,7 @@ def post_tts(
     ref_audio: Path | None,
     ref_text: str | None,
     timeout: int,
+    language: str = "english",
 ) -> bytes:
     """POST multipart by hand; the boundary must match byte for byte."""
     boundary = f"----umbra{uuid.uuid4().hex}"
@@ -123,6 +131,9 @@ def post_tts(
         parts.append(b"\r\n")
 
     field("text", text)
+    # Must be sent: the server used to hardcode language="english" in every
+    # branch, so German and French were synthesised with English phonology.
+    field("language", language)
     if prompt_slug:
         field("prompt_slug", prompt_slug)
     if ref_audio:
@@ -286,6 +297,7 @@ def main() -> int:
                     ref_audio=None if use_prompt else refs[slug],
                     ref_text=None if use_prompt else REF_TEXT,
                     timeout=args.timeout,
+                    language=LANG_CODES.get(suffix, "english"),
                 )
                 dest.write_bytes(audio)
                 made.append(dest)