marksverdhei/Qwen3-Voice-Embedding-12Hz-1.7B-onnx

Qwen3-Voice-Embedding-12Hz-1.7B (ONNX)

ONNX exports of the Qwen3-Voice-Embedding-12Hz-1.7B ECAPA-TDNN speaker encoder. Produces 2048-dimensional x-vector speaker embeddings from audio.

Three quantization variants are provided for different deployment targets:

Usage

Python (ONNX Runtime)

import numpy as np
import onnxruntime as ort
import librosa

# Load model
session = ort.InferenceSession("speaker_encoder_fp32.onnx")

# Compute mel spectrogram (must match training preprocessing)
audio, sr = librosa.load("audio.wav", sr=24000, mono=True)
mel = librosa.feature.melspectrogram(
    y=audio, sr=24000, n_fft=1024, hop_length=256,
    n_mels=128, fmin=0, fmax=12000,
)
mel = np.log(np.clip(mel, a_min=1e-5, a_max=None))
mel = mel.T[np.newaxis, ...]  # (1, time, 128)

# Run inference
embedding = session.run(None, {"mel_spectrogram": mel.astype(np.float32)})[0]
# embedding shape: (1, 2048)

Browser (ONNX Runtime Web)

import * as ort from "onnxruntime-web";

const session = await ort.InferenceSession.create("speaker_encoder_fp16.onnx");

// mel: Float32Array of shape [1, time_steps, 128]
const tensor = new ort.Tensor("float32", mel, [1, timeSteps, 128]);
const results = await session.run({ mel_spectrogram: tensor });
const embedding = results.speaker_embedding.data; // Float32Array(2048)

Input / Output

The time axis is dynamic — any length mel spectrogram is accepted.

Audio Preprocessing

Export Details

Exported with torch.onnx.export (opset 18) from the standalone PyTorch model. Verified against PyTorch output (cosine similarity = 1.0, max abs diff < 3e-6).

Related Models

• marksverdhei/Qwen3-Voice-Embedding-12Hz-1.7B — PyTorch source model
• marksverdhei/Qwen3-Voice-Embedding-12Hz-0.6B — 1024-dim variant (PyTorch)
• marksverdhei/Qwen3-Voice-Embedding-12Hz-0.6B-onnx — 1024-dim variant (ONNX)

Citation

@article{qwen3-tts,
  title={Qwen3-TTS Technical Report},
  author={Hu, Hangrui and Zhu, Xinfa and He, Ting and Guo, Dake and Zhang, Bin and Wang, Xiong and Guo, Zhifang and Jiang, Ziyue and Hao, Hongkun and Guo, Zishan and Zhang, Xinyu and Zhang, Pei and Yang, Baosong and Xu, Jin and Zhou, Jingren and Lin, Junyang},
  journal={arXiv preprint arXiv:2601.15621},
  year={2026}
}

@article{ecapa-tdnn,
  title={ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification},
  author={Desplanques, Brecht and Thienpondt, Jenthe and Demuynck, Kris},
  journal={Proc. Interspeech},
  year={2020}
}