SPRINGLab/SPRING_F5

Cross-Lingual TTS for Indian Languages: A Benchmark and Inference-Time Duration Estimation Strategies

SPRING_F5 is a multilingual text-to-speech (TTS) model based on F5-TTS, fine-tuned to support 23 Indian Language & English.

Supported Languages

SPRING_F5 supports the following 24 languages:

• Assamese, Bengali, Bodo, Gujarati, Hindi, Kannada, Kashmiri, Konkani, Maithili, Malayalam, Manipuri, Marathi, Nepali, Odia, Punjabi, Sanskrit, Santali, Sindhi, Tamil, Telugu, Urdu, Dogri, Rajasthani and English

🚀 Installation

We recommend using a dedicated Conda environment.

conda create -n springf5 python=3.10 -y
conda activate springf5
pip install git+https://github.com/ArigalaAdarsh/SPRING_F5.git

from transformers import AutoModel
import numpy as np
import soundfile as sf

# Load INF5 from Hugging Face
repo_id = "SPRINGLab/SPRING_F5"
model = AutoModel.from_pretrained(repo_id, trust_remote_code=True)

# Generate speech
audio = model(" 2026 లో ఈ project లో మా ultimate goal ఏంటంటే, ఒక Telugu speaker తన daily conversation లో naturally English words ఉపయోగించినప్పుడు, మన T T S system కూడా అదే style ని understand చేసి, unnecessary pauses లేకుండా చాలా smooth మరియు natural speech ని generate చేయాలి.", 
            ref_audio_path="prompt_audios/example1_te.wav",
            ref_text="వందేభారత్ రైలు ఆధునిక భారతదేశం యొక్క వైభవోపేతమైన చిత్రాలలో ఒకటిగా ఉంది",
            lang='te'  # Language ID is used for number-to-Indic word conversion.
                  )

# Normalize and save output
if audio.dtype == np.int16:
    audio = audio.astype(np.float32) / 32768.0
sf.write("generated_audios/outputs/example.wav", np.array(audio, dtype=np.float32), samplerate=24000)

Training Details

SPRING_F5 was trained using 2 × NVIDIA H200 GPUs for approximately two weeks.

We would like to sincerely thank C-DAC for providing the computational resources required to train this model.

We used the F5-TTS Base configuration, with approximately 330M parameters, following the model configuration described in the original F5-TTS work.

Training Data

The model was trained on approximately 3,220 hours of high-quality speech data collected from the following datasets:

• Rasa
• IndicTTS
• IndicVoices-R

These datasets provide diverse multilingual speech data covering the Indian languages supported by SPRING_F5.

---

References

We would like to express our sincere gratitude to the authors and contributors of F5-TTS for their valuable contributions to text-to-speech research and for providing the foundation for this work.

SPRING_F5 builds upon the F5-TTS architecture and training methodology and extends it toward multilingual speech synthesis for Indian languages.

@inproceedings{arigala-etal-2026-cross,
    title = "Cross-Lingual {TTS} for {I}ndian Languages: A Benchmark and Inference-Time Duration Estimation Strategies",
    author = "Arigala, Adarsh  and
      Gangwar, Arjun  and
      Umesh, Srinivasan",
    editor = "Wang, Xiaoman  and
      Wang, Qianyu  and
      Liu, Zhengyuan  and
      Chen, Nancy F.  and
      Yuan, Zheng  and
      Lan, Yunshi  and
      Jiang, Jing",
    booktitle = "Proceedings of the 1st Workshop on Multi-modality and Multi-lingual Language Learning",
    month = nov,
    year = "2026",
    address = "Hengqin, Zhuhai, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.multilll-1.6/",
    doi = "10.18653/2026.multilll-1.6",
    pages = "47--55",
    ISBN = "979-8-89176-449-1"
}