tv-orpheus-v2

Thorsten-Voice – Orpheus TTS v2 (Mini Fine-Tuned)

Overview

Thorsten-Voice/tv-orpheus-v2 is an improved version of tv-orpheus-v1, further optimized to better match the natural speaking style of the original speaker.

It was fine-tuned using a small, carefully curated mini dataset (60 recordings, TV-24kHz-2025.12-Neutral-FT-Mini) recorded in everyday speech situations, focusing on:

• calm explanations
• natural pauses
• neutral statements
• questions
• short and long sentences

Training date / codebase reference: December 2025

---

Training Data

Base Training (v1)

• ~12,000 recordings
• Thorsten-Voice Dataset 2022.10
• Sample rate: 24 kHz
• License: CC0

Additional Fine-Tuning (v2)

• Dataset: Thorsten-Voice/TV-24kHz-2025.12-Neutral-FT-Mini
• Number of recordings: 60
• Sample rate: 24 kHz
• Content: neutral, natural everyday speech
• License: CC0 (Public Domain)

The mini dataset was designed to pull the synthesized voice closer to the real speaker, improving authenticity and reducing synthetic artifacts.

---

Model Lineage

• Base model: canopylabs/3b-de-ft-research_release
• Intermediate model: Thorsten-Voice/tv-orpheus-v1
• Final fine-tuning: 60-sample neutral mini dataset
• TTS framework: Orpheus TTS

---

License Clarification

• Model weights:
  Licensed under Apache-2.0, inherited from the Orpheus base model.

• Training data & voice:
  Released under CC0 (Public Domain) by the speaker.

No additional usage restrictions are imposed.

---

Usage Example

python inference_hf.py \
  --model_path Thorsten-Voice/tv-orpheus-v2 \
  --text "Für mich sind alle Menschen gleich, unabhängig von Herkunft oder Religion." \
  --outfile output.wav

TTS generation script (inference_hf.py) is included in repo (see files) for easier generation of Thorsten-Voice powered by OrpheusTTS.

Notes on Voice Quality

Compared to v1, this model:

• sounds closer to the real speaker
• uses a more natural rhythm
• shows improved intonation in longer sentences
• Minor artifacts such as occasional unexpected pauses may still occur and are subject to future refinements.

Acknowledgements

Special thanks to Orpheus TTS and its authors for providing an open and high-quality German TTS foundation.

• Orpheus TTS GitHub: https://github.com/canopylabs/orpheus-tts
• Base model by: Canopy Labs

Thorsten-Voice Project

Thorsten-Voice is an open voice project aimed at providing freely usable German speech data and models.

• Project page: https://www.Thorsten-Voice.de