---
language:
- en
tags:
- speech
- spoken-language-model
- full-duplex
- voice-agent
- taste
- text-to-speech
license: other
license_name: taste2-research-only
license_link: https://gitycc.github.io/TASTE2-Homepage/
---

# TASTE2-8B-EN

Part of the **TASTE2** project — *Text-Aligned Speech Modeling and Deployment toward Full-Duplex Voice Interaction*.

- 📄 Paper: [TASTE2 paper (PDF)](https://gitycc.github.io/TASTE2-Homepage/assets/taste2-paper.pdf)
- 🌐 Project homepage: https://gitycc.github.io/TASTE2-Homepage/
- 💻 Spoken LM code: https://github.com/GitYCC/TASTE-SpokenLM-2
- 💻 VoiceBot code: https://github.com/GitYCC/TASTE-Voice-Bot
- 🤗 All TASTE2 checkpoints: https://huggingface.co/collections/YC-Chen/taste2
- 📚 Dataset: https://huggingface.co/datasets/wilzzzz/paralinguistic_dialogues

## About TASTE2

Full-duplex voice interaction requires more than converting one complete utterance into another. A system must process speech as it arrives, decide when to take or yield the floor, and stop when the user interrupts, while preserving pretrained linguistic competence and acoustic paralinguistic cues. TASTE2 turns the original utterance-level TASTE method into an incremental dialogue stack of three components — a **Speech Tokenizer**, a **Spoken LM**, and a **Speech Detokenizer** — aligning one continuous audio latent with every text token so the token sequence length is preserved end-to-end.

A shared text-token vocabulary removes word-level averaging and language-dependent segmentation, modality-aligned dialogue training predicts one continuous audio latent per text token without interleaving heterogeneous token streams, and an incremental Speech Detokenizer streams synthesis through CosyVoice2. On top of this, **TASTE2 VoiceBot** is a working research system that processes user speech incrementally, streams synthesized audio, and stops generation when the user barges in.

## This checkpoint

`TASTE2-8B-EN` is a base TASTE2 checkpoint for **English**, built on the **Qwen2-7B** backbone and trained through the TASTE2 stage-1 and stage-2 training stages (`taste2_stage1.yaml`, `taste2_stage2.yaml`). It is not instruction-tuned or dialogue-fine-tuned — for instruction-following / dialogue-fine-tuned variants, see the other checkpoints in the [TASTE2 collection](https://huggingface.co/collections/YC-Chen/taste2), such as [`TASTE2-8B-EN-Instruct-merge-FT`](https://huggingface.co/YC-Chen/TASTE2-8B-EN-Instruct-merge-FT) or [`TASTE2-8B-EN-Instruct-dircect-FT`](https://huggingface.co/YC-Chen/TASTE2-8B-EN-Instruct-dircect-FT).

## Files

This repo bundles the full TASTE2 inference stack for this checkpoint: Spoken LM weights (`llm.pt`, `slm.pt`), CosyVoice2-based flow/vocoder weights (`flow.pt`, `flow.cache.pt`, `hift.pt`, `flow.encoder.fp16.zip`, `flow.encoder.fp32.zip`, `flow.decoder.estimator.fp32.onnx`), the speech tokenizer (`speech_tokenizer_v2.onnx`, `campplus.onnx`), the bundled `qwen2-7b/` and `CosyVoice-BlankEN/` base checkpoints, `distil-whisper/`, and the stage-1/stage-2 training config yamls.

For setup and inference instructions, see the [TASTE-SpokenLM-2](https://github.com/GitYCC/TASTE-SpokenLM-2) and [TASTE-Voice-Bot](https://github.com/GitYCC/TASTE-Voice-Bot) repositories.

## Authors

Yi-Chang Chen*<sup>1</sup>, Chun Wei Chen*<sup>2,3</sup>, Dien-Ruei Wu*<sup>2,3</sup>, Jie Lin<sup>3</sup>, Hung-yi Lee<sup>3</sup>, Da-Shan Shiu<sup>1</sup>

<sup>1</sup> MediaTek Research · <sup>2</sup> Internship at MediaTek Research · <sup>3</sup> National Taiwan University · *Equal contribution

## Citation

BibTeX will be added once the public paper record is available. In the meantime, please refer to the [project homepage](https://gitycc.github.io/TASTE2-Homepage/) and the [paper PDF](https://gitycc.github.io/TASTE2-Homepage/assets/taste2-paper.pdf).

## License / Disclaimer

This checkpoint is provided for **academic and research purposes only**, © 2026 MediaTek Research and National Taiwan University. It is built on top of `Qwen2-7B` and `CosyVoice2`, which retain their own respective upstream licenses — please review those separately before use.