BananaMind-TTS-V2.1-Preview

BananaMind TTS V2.1 Preview

BananaMind TTS V2.1 Preview, codename Mica, is a fixed-voice, single-speaker text-to-speech release with separate English and German locale folders.

Locale Folders

• en-us: English acoustic model and vocoder from BananaMind TTS V2.
• de-de: German acoustic model trained on the ThorstenVoice Dataset 2022.10 and a German HiFi-GAN vocoder.

Each locale folder contains:

• model.safetensors: FP32 Tacotron-lite acoustic model
• vocoder.safetensors: default BF16 HiFi-GAN generator-only vocoder
• FP32/vocoder.safetensors: FP32 HiFi-GAN generator-only vocoder
• full_vocoder.pt: full HiFi-GAN training checkpoint for resume/fine-tuning
• config.json: locale config for Transformers remote code
• model_config.json: sidecar metadata

The full_vocoder.pt files include generator, discriminators, optimizer states, config, epoch, and step. They are not needed for normal inference; use vocoder.safetensors or FP32/vocoder.safetensors for generation.

LJ Speech Attribution

The English en-us model was trained using The LJ Speech Dataset.

BibTeX:

@misc{ljspeech17,
  author       = {Keith Ito and Linda Johnson},
  title        = {The LJ Speech Dataset},
  howpublished = {\url{https://keithito.com/LJ-Speech-Dataset/}},
  year         = 2017
}

Usage

Install runtime dependencies:

pip install torch numpy safetensors transformers huggingface_hub

Generate English:

python generate.py --language en-us --text "This is BananaMind TTS version two point one."

Generate German:

python generate.py --language de-de --text "Hallo, das ist BananaMind TTS Version zwei Punkt eins."

You can also load a locale folder directly:

import torch
from transformers import AutoModel

model = AutoModel.from_pretrained("BananaMind-TTS-V2.1/de-de", trust_remote_code=True)
model.eval()

with torch.inference_mode():
    out = model.tts("Hallo, das ist ein deutscher Sprachtest.", normalize_wav=True)

model.save_wav("sample_de.wav", out.waveform, out.sample_rate)

Thorsten Voice Attribution

The German de-de model was trained using the ThorstenVoice Dataset 2022.10.

Required dataset citation/credit:

• ThorstenVoice Dataset 2022.10
• Voice: Thorsten Müller
• Audio optimization: Dominik Kreutz
• DOI: 10.5281/zenodo.7265581
• License: CC0
• Project: https://www.thorsten-voice.de/

The ThorstenVoice dataset README says: "If you use this Thorsten voice dataset please quote it using Zendodo DOI: 10.5281/zenodo.7265581".

BibTeX:

@misc{Muller_Thorsten-Voice,
author = {Müller, Thorsten and Kreutz, Dominik},
title = {{Thorsten-Voice}},
url = {https://github.com/thorstenMueller/Thorsten-Voice}
}

Limitations

• Fixed voices only; this is not voice cloning.
• No speaker embeddings or reference audio conditioning.
• Character tokenizer; German umlauts are supported in the de-de tokenizer, but this is not a phoneme model.
• Numbers should be written as words for best results.