Qwen3 Forced Aligner ONNX

This repository contains a dynamic ONNX export of Qwen/Qwen3-ForcedAligner-0.6B for forced alignment.

It preserves the original tokenizer, processor, and configuration files while replacing the PyTorch weights with an ONNX Runtime graph that supports:

dynamic batch size
dynamic text sequence length
dynamic audio feature length
browser loading with transformers.js v4

Base model

This is a format conversion of Qwen/Qwen3-ForcedAligner-0.6B. It does not introduce new training or fine-tuning.

Files

onnx/model.onnx
onnx/model.onnx_data
onnx/model_q4.onnx
config.json
generation_config.json
tokenizer and processor files
export_metadata.json

Usage

Python (`qwen_asr` ONNX backend)

from qwen_asr import Qwen3ForcedAligner

aligner = Qwen3ForcedAligner.from_pretrained(
    "valoomba/Qwen3-ForcedAligner-0.6B-ONNX",
    backend="onnx",
)

Browser / Node (`transformers.js` v4)

Use AutoTokenizer, AutoFeatureExtractor, and AutoModel with the repo root. The ONNX graph lives at onnx/model.onnx, and config.json includes transformers.js_config.use_external_data_format for the external data sidecar.

For browser-friendly loading, this repo also includes a q4 variant at onnx/model_q4.onnx.

const model = await AutoModel.from_pretrained("valoomba/Qwen3-ForcedAligner-0.6B-ONNX", { dtype: "q4" });

Export provenance

Source model: Qwen/Qwen3-ForcedAligner-0.6B
Generated at: 2026-03-31 09:04:54 UTC
Export tool: examples/export_qwen3_forced_aligner_onnx.py

valoomba/Qwen3-ForcedAligner-0.6B-ONNX