Moonshine ASR pipeline throws 'Unable to decode without a tokenizer' — _call_moonshine calls processor.batch_decode but processor has no tokenizer

[mosabsayyed]

Bug

The automatic-speech-recognition pipeline crashes at decode time for all moonshine models (e.g. onnx-community/moonshine-tiny-ar-ONNX, onnx-community/moonshine-tiny-en-ONNX), in both 3.8.1 and 4.2.0.

Exact error

Error: Unable to decode without a tokenizer.
    at Function.batch_decode (transformers.node.mjs:15763)
    at Function._call_moonshine (transformers.node.mjs:32216)

Reproduce (Node, v4.2.0)

import { pipeline, env } from "@huggingface/transformers";
env.allowLocalModels = false;
const transcriber = await pipeline("automatic-speech-recognition", "onnx-community/moonshine-tiny-ar-ONNX", { dtype: "q8", device: "cpu" });
// model loads fine (~3s)
const out = await transcriber(float32Audio16k); // throws at decode

Root cause

_call_moonshine calls the processor's batch_decode:

// src/pipelines/automatic-speech-recognition.js:328 (v4.2.0)
const text = this.processor.batch_decode(outputs, { skip_special_tokens: true })[0];

But the moonshine processor is constructed without a tokenizer — only a feature_extractor. Verified directly:

const proc = await AutoProcessor.from_pretrained("onnx-community/moonshine-tiny-ar-ONNX");
// proc.components => ['feature_extractor']   (no 'tokenizer')
// proc.config.processor_class => undefined
// proc.config.auto_map => undefined

So processor.batch_decode hits its if (!this.tokenizer) throw guard (processing_utils.js:94).

The pipeline-level tokenizer is correctly populated:

const p = await pipeline("automatic-speech-recognition", "onnx-community/moonshine-tiny-ar-ONNX", {...});
// p.tokenizer => PreTrainedTokenizer  (loaded from tokenizer.json) ✓
// p.processor.tokenizer => undefined   ✗

Fix

_call_moonshine should use the pipeline's tokenizer, not the processor's:

const text = this.tokenizer.batch_decode(outputs, { skip_special_tokens: true })[0];

This matches every other ASR path — _call_whisper uses this.tokenizer._decode_asr, and _call_wav2vec2 uses this.tokenizer.decode. _call_moonshine is the only one that routes through this.processor, and the processor legitimately may not carry a tokenizer.

Environment

• @huggingface/transformers: 3.8.1 (reproduced) and 4.2.0 (reproduced)
• Model: onnx-community/moonshine-tiny-ar-ONNX (and the en variant)
• Runtime: Node 22 (onnxruntime-node cpu) and browser (onnxruntime-web wasm)

Thank you.