Helsinki-NLP published the English→Japanese model as opus-mt-en-jap - using jap, not the standard code ja - while the reverse direction (opus-mt-ja-en) correctly uses ja. That's an inconsistency in their own catalog, not my logic, but my code assumed standard ISO codes everywhere, so it was 404ing on a model that actually exists.

app.py

@@ -325,7 +325,7 @@ with gr.Blocks(title="EchoScript") as demo:
 **Upload Audio → Select Audio Window → Detect Language & Generate Transcript
 → Preview & Choose Languages → Generate Translations → Copy / Download**
 
-<sub>build: 2026-06-27 00:24 UTC &middot; real per-source-language Marian availability (no more fixed 4-language guess)</sub>
+<sub>build: 2026-06-28 22:17 UTC &middot; fixed en-jap Helsinki-NLP naming exception</sub>
         """
     )
 

services/translation.py

@@ -24,12 +24,13 @@ whether an Anthropic API key was supplied -- never stored:
   through to the Anthropic client for that one call and is never written
   to disk, logged, or cached in any module-level state.
 - "marian": fully offline, no API key needed. Uses local Helsinki-NLP
-  MarianMT models via `transformers`, restricted to the small set of
-  languages where direct or English-pivoted coverage is reliable.
+  MarianMT models via `transformers`. Coverage is checked for real, per
+  source language, against the Hub (direct pair or English pivot) -- see
+  available_marian_targets() -- rather than assumed from a fixed list.
 
 The UI is expected to only offer the larger ANTHROPIC_TARGET_LANGUAGES
-list once a key has been entered, and MARIAN_TARGET_LANGUAGES otherwise --
-see app.py.
+list once a key has been entered, and available_marian_targets(source)
+otherwise -- see app.py.
 """
 
 from __future__ import annotations
@@ -195,7 +196,7 @@ def _try_load_marian_engine(source_language: str, target_language: str):
     """
     from transformers import AutoModelForSeq2SeqLM, AutoTokenizer  # heavy import, deferred
 
-    model_name = f"Helsinki-NLP/opus-mt-{source_language}-{target_language}"
+    model_name = _marian_repo_id(source_language, target_language)
     try:
         tokenizer = AutoTokenizer.from_pretrained(model_name)
         model = AutoModelForSeq2SeqLM.from_pretrained(model_name)
@@ -204,6 +205,23 @@ def _try_load_marian_engine(source_language: str, target_language: str):
     return tokenizer, model
 
 
+# Helsinki-NLP doesn't always use standard ISO 639-1 codes in its repo
+# names. Known exceptions go here, mapped to the actual code used in that
+# specific repo name -- e.g. the en->Japanese model is published as
+# "Helsinki-NLP/opus-mt-en-jap" (not "...-en-ja"), even though the reverse
+# direction "Helsinki-NLP/opus-mt-ja-en" correctly uses "ja". Without this,
+# every availability check for Japanese-as-a-target would 404 on a model
+# that actually exists, regardless of source language.
+_MARIAN_CODE_ALIASES: dict[tuple[str, str], str] = {
+    ("en", "ja"): "jap",
+}
+
+
+def _marian_repo_id(source_language: str, target_language: str) -> str:
+    aliased_target = _MARIAN_CODE_ALIASES.get((source_language, target_language), target_language)
+    return f"Helsinki-NLP/opus-mt-{source_language}-{aliased_target}"
+
+
 @lru_cache(maxsize=None)
 def _marian_model_exists(source_language: str, target_language: str) -> bool:
     """Check (and cache) whether Helsinki-NLP publishes this exact pair.
@@ -222,7 +240,7 @@ def _marian_model_exists(source_language: str, target_language: str) -> bool:
     from huggingface_hub.utils import HfHubHTTPError
 
     try:
-        HfApi().model_info(f"Helsinki-NLP/opus-mt-{source_language}-{target_language}")
+        HfApi().model_info(_marian_repo_id(source_language, target_language))
         return True
     except HfHubHTTPError:
         return False