README.md

---
base_model:
- canopylabs/orpheus-3b-0.1-ft
tags:
- transformers
- unsloth
- tts
- orpheus
license: apache-2.0
language:
- en
datasets:
- innovationm-ai/dailytalk-male
- innovationm-ai/dailytalk-female
pipeline_tag: text-to-speech
library_name: transformers
---

<div align="center">
  <img src="./logo.png" alt="InnovationM"/>
  <br/><br/>

  <a href="https://www.innovationm.com/">
    <img src="https://img.shields.io/badge/InnovationM-Website-orange?style=for-the-badge&logo=google-chrome&logoColor=white" alt="InnovationM Website Badge"/>
  </a>
  <a href="https://in.linkedin.com/company/innovationm">
    <img src="https://img.shields.io/badge/LinkedIn-Profile-blue?style=for-the-badge&logo=linkedin&logoColor=white" alt="LinkedIn Badge"/>
  </a>
  <br/><br/>

  <a href="https://colab.research.google.com/drive/17ZeVLF8Q1P9juIKftijN9uwl2LrfMeCz?usp=drive_link">
    📓 Colab Notebook: Orpheus 3B TTS Finetuning
  </a>
</div>


# **Orpheus-3B (Fine-Tuned) — InnovationM**

This repository contains a fine-tuned variant of **Orpheus TTS 3B** — a state-of-the-art speech generation framework originally developed by [Canopy Labs](https://canopylabs.ai/).  

# **Fine-Tuned Model Description**

This model is derived from the base checkpoint **`canopylabs/orpheus-3b-0.1-ft`**, fine-tuned using the **Unsloth optimization framework**, which enables up to **2× faster training** with reduced memory footprint.

Additional alignment and adaptation were performed with the **Hugging Face TRL** (Training Reinforcement Learning) library.

### **Audio Samples**

#### **Before Finetuning (Female Voice)**

<audio controls>
  <source src="https://huggingface.co/aiAtInnovationM/orpheus-tts-3b-finetuned/resolve/main/tts_before_tuning_1.wav" type="audio/wav">
  Your browser does not support the audio element.
</audio>

#### **After Finetuning (Male Voice)**

<audio controls>
  <source src="https://huggingface.co/aiAtInnovationM/orpheus-tts-3b-finetuned/resolve/main/tts_after_tuning_1.wav" type="audio/wav">
  Your browser does not support the audio element.
</audio>

### **Fine-Tuned Model Information**
- **Fine-tuned by:** `InnovationM's AI Team`  
- **Organization:** [InnovationM](https://www.innovationm.com/)  
- **License:** Apache-2.0  
- **Base Model:** `canopylabs/orpheus-3b-0.1-ft`  
- **Dataset Used:** `aiAtInnovationM/dailytalk-male`

# **Special Thanks**

- **Canopy Labs Research Team** — for creating and open-sourcing the original CSM architecture and research.  
- **Unsloth Team** — for providing the high-efficiency training framework enabling optimized fine-tuning workflows.  
- **Meta AI & Llama Team** — for developing and releasing the foundational Llama models that support the Orpheus architecture and its derivatives.

# **Authors**

### **Original Model Authors**  
**Canopy Labs** Team.

### **Fine-Tuned Model Authors**  
The **InnovationM** team.

# 📎 **Useful Resources**

| Resource | Link |
|---------|------|
| InnovationM Website | https://www.innovationm.com/ |
| InnovationM LinkedIn | https://in.linkedin.com/company/innovationm |
| Fine-Tuning Notebook | https://colab.research.google.com/drive/17ZeVLF8Q1P9juIKftijN9uwl2LrfMeCz?usp=drive_link |
| Base Model | https://huggingface.co/unsloth/csm-1b |

# 📄 **Citation**

If you use this model in your research, please cite both **Canopy Labs Model** and this **InnovationM fine-tuned version**.

```bibtex
@misc{innovationm2025,
  title = {Orpheus-3B Fine-Tuned Model},
  author = {InnovationM},
  howpublished = {Hugging Face},
  year = {2025}
}