From 56a9de4278cc9801afc590b45961243dadd80362 Mon Sep 17 00:00:00 2001
From: jeevan6996 <jeevanpawar5890@gmail.com>
Date: Fri, 14 Aug 2026 19:08:20 +0100
Subject: [PATCH] Fix Moonshine ASR tokenizer decoding

---
 .../pipelines/automatic-speech-recognition.js |  2 +-
 ..._pipelines_automatic_speech_recognition.js | 27 +++++++++++++++++++
 2 files changed, 28 insertions(+), 1 deletion(-)

diff --git a/packages/transformers/src/pipelines/automatic-speech-recognition.js b/packages/transformers/src/pipelines/automatic-speech-recognition.js
index 238bb45a4..429303f7c 100644
--- a/packages/transformers/src/pipelines/automatic-speech-recognition.js
+++ b/packages/transformers/src/pipelines/automatic-speech-recognition.js
@@ -325,7 +325,7 @@ export class AutomaticSpeechRecognitionPipeline
             const max_new_tokens = Math.floor(aud.length / sampling_rate) * 6;
             const outputs = await this.model.generate({ max_new_tokens, ...kwargs, ...inputs });
 
-            const text = this.processor.batch_decode(/** @type {Tensor} */ (outputs), { skip_special_tokens: true })[0];
+            const text = this.tokenizer.batch_decode(/** @type {Tensor} */ (outputs), { skip_special_tokens: true })[0];
             toReturn.push({ text });
         }
         return single ? toReturn[0] : toReturn;
diff --git a/packages/transformers/tests/pipelines/test_pipelines_automatic_speech_recognition.js b/packages/transformers/tests/pipelines/test_pipelines_automatic_speech_recognition.js
index 33d67ac92..626b9503f 100644
--- a/packages/transformers/tests/pipelines/test_pipelines_automatic_speech_recognition.js
+++ b/packages/transformers/tests/pipelines/test_pipelines_automatic_speech_recognition.js
@@ -7,6 +7,33 @@ const PIPELINE_ID = "automatic-speech-recognition";
 
 export default () => {
   describe("Automatic Speech Recognition", () => {
+    it("uses the pipeline tokenizer to decode Moonshine output", async () => {
+      const model = {
+        config: { model_type: "moonshine" },
+        generate: async () => ({}),
+        dispose: async () => {},
+      };
+      const processor = Object.assign(async () => ({}), {
+        feature_extractor: { config: { sampling_rate: 16000 } },
+      });
+      let decodeCalls = 0;
+      const tokenizer = {
+        batch_decode: () => {
+          decodeCalls += 1;
+          return ["decoded text"];
+        },
+      };
+      const pipe = new AutomaticSpeechRecognitionPipeline({
+        task: PIPELINE_ID,
+        model,
+        tokenizer,
+        processor,
+      });
+
+      await expect(pipe._call_moonshine(new Float32Array(16000), {})).resolves.toEqual({ text: "decoded text" });
+      expect(decodeCalls).toBe(1);
+    });
+
     describe("whisper (tiny-random)", () => {
       const model_id = "Xenova/tiny-random-WhisperForConditionalGeneration";
       const SAMPLING_RATE = 16000;