moonshine stream: cap the final decode budget at the model card's output length

CHANGELOG.md

@@ -74,6 +74,24 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   prefill is itself out-of-distribution for it). Measured on an 8-utterance zero-shot tool-calling set: 0/8 and 1/8
   under generic ChatML / `QwenChatTemplate` vs 4-5/8 with this template.
 
+### Changed — Moonshine v2 streaming: the final decode budget follows the model card (#455)
+
+- **`nativeFinish` caps the exact re-decode at `max_new_tokens = samples * 6.5/16000 + 2`**
+  (`llm-runtime:iree-android`, `moonshine_stream_jni.c`), the output-length cap the Moonshine v2
+  model card specifies, instead of a fixed 24 tokens regardless of how much audio the utterance
+  held. Without it a short clip keeps decoding long after the audio is spent, and the greedy decode
+  spends the remaining budget restarting the utterance rather than stopping. The budget is floored
+  at 4 so a one-word command still has room and ceilinged at the previous fixed 24, so nothing
+  decodes longer than before — above roughly 3.4 s of audio the formula exceeds the ceiling and the
+  change is a no-op. The `finish` trace line now reports tokens used against the budget.
+- **Measured** on 183 German command-and-control recordings on a Mali device, using clips above the
+  threshold as a control group: below it the turn finishes **0.58 s sooner** (paired median, faster
+  in 135 of 172); above it unchanged (+0.09 s, 11 recordings). Median word error rate does not move.
+  The saving appears exactly where the cap can act and nowhere else.
+- The zero-padded final encoder window is a separate, still-open gap (#458): the encoder graph takes
+  features only and carries no attention mask, and moving the window is not the fix — a variant that
+  end-aligned it measured no effect on those same 11 control recordings and was dropped from #455.
+
 ## [0.56.2] — 2026-09-22
 
 A transformers-only release against **SKaiNET engine 0.56.0** (unchanged).

llm-runtime/iree-android/native/moonshine_stream_jni.c

@@ -83,8 +83,8 @@
 #define FE_OUT (CHUNK + FE_LC)             /* 68 frames out */
 #define MAXMEM 256        /* compiled cross-memory pad (5.12 s of finalized speech) */
 #define MAXTOK 48
-#define FINISH_TOKENS 24  /* C&C final budget: 5.12 s of speech never needs more; halves the
-                             worst-case finalize when the decode loops instead of stopping */
+#define FINISH_TOKENS 24  /* hard ceiling for the final budget (see finish_budget) */
+#define TOKENS_PER_SAMPLE (6.5f / 16000.0f)  /* the model card's cap: max_new_tokens = samples * 6.5/16000 + 2 */
 #define HOP_TOKENS 6      /* max new tokens decoded per hop (budget guard) */
 #define RESTART_HOPS 2    /* first hops re-decode exactly instead of incrementally */
 #define PEEK_FRAMES 36    /* early-peek window: first partial at ~0.72 s instead of 1.28 s */
@@ -273,6 +273,22 @@ static iree_hal_buffer_view_t* run1(Engine* e, iree_runtime_session_t* s, const
   return out;
 }
 
+/* The model card's output-length cap, in tokens, for `npcm` samples of audio:
+ *   max_new_tokens = samples * 6.5 / 16000 + 2
+ * A fixed budget lets a short clip keep decoding long after the audio is spent, and the greedy
+ * decode spends that budget restarting the utterance rather than stopping. Floored at 4 so a
+ * one-word command still has room, ceilinged at the previous fixed budget so nothing decodes
+ * longer than before — above ~3.4 s of audio the formula exceeds the ceiling and this is a no-op.
+ * Measured on 183 German command-and-control recordings on a Mali device: clips below that
+ * threshold finish 0.58 s sooner (paired median, faster in 135 of 172), clips above it are
+ * unchanged (+0.09 s, 11 recordings), and the median word error rate does not move. */
+static int finish_budget(int npcm) {
+  int b = (int)(npcm * TOKENS_PER_SAMPLE) + 2;
+  if (b < 4) b = 4;
+  if (b > FINISH_TOKENS) b = FINISH_TOKENS;
+  return b;
+}
+
 /* Process one encoder window starting at feature frame `start`; finalize rows into e->mem. */
 static int process_window(Engine* e, int start, int produced, int flush, int peek) {
   double t0 = now_ms();
@@ -706,7 +722,7 @@ JNIEXPORT jstring JNICALL JNIFN(nativeFinish)(JNIEnv* env, jobject thiz, jlong h
   if (e->nmem > 0) {
     /* exact: fresh decode over the final memory */
     release_self(e); e->ntoks = 0; e->pos = 0; e->text[0] = 0;
-    if (run_prefill(e, 1) == 0) run_steps(e, FINISH_TOKENS, 1);
+    if (run_prefill(e, 1) == 0) run_steps(e, finish_budget(e->npcm), 1);
     if (!e->ended_eos) drop_restart_suffix(e);
     /* C&C utterances are single sentences; on a silent tail the greedy decode restarts the
      * utterance instead of emitting EOS ("Go to the next. Go to next") — same failure and
@@ -716,8 +732,8 @@ JNIEXPORT jstring JNICALL JNIFN(nativeFinish)(JNIEnv* env, jobject thiz, jlong h
     }
     const char* p = e->text; while (*p == ' ') ++p;
     out = (*env)->NewStringUTF(env, p);
-    TLOG("finish %.0fms nmem %d toks %d \"%s\"\n",
-            now_ms() - t0, e->nmem, e->ntoks, p);
+    TLOG("finish %.0fms nmem %d toks %d/%d \"%s\"\n",
+            now_ms() - t0, e->nmem, e->ntoks, finish_budget(e->npcm), p);
   }
   reset_utterance(e);
   return out;