This model is a fine-tuned LoRA adapter designed to adapt the Orpheus base model for Text-to-Speech (TTS) synthesis in Swiss German. It processes text conditioning prompts and generates speech tokens encoded via a 24kHz SNAC codec interleaving layout.

Model Details

Model Description

Developed by: griffing52
Shared by: griffing52
Model type: PEFT Adapter (LoRA) for Causal Language Modeling (Speech/Text generation)
Language(s) (NLP): Swiss German (gsw), German (de)
License: [More Information Needed]
Finetuned from model: canopylabs/3b-de-ft-research_release

Model Sources

Repository: griffing52/orpheus-swiss-german-lora

Uses

Direct Use

This adapter is intended for Text-to-Speech generation in Swiss German dialects. Speech synthesis is triggered using a formatted prefix prompt layout: voice: transcript.

Out-of-Scope Use

This model is not built for standard text-based conversational modeling, nor will it perform well generating languages or accents far outside German/Swiss German linguistic structures.

Bias, Risks, and Limitations

Dialect Consistency: Swiss German features massive lexical and phonetic variance across regions. Dialect spelling inconsistencies in training transcripts heavily influence the resulting prosody and pronunciation quality.
Overtraining Risk: Training beyond 1–2 epochs on small datasets risks degrading prosody and speech naturalness.

How to Get Started with the Model

You can load this adapter on top of the base model using the code below:

python

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE_MODEL = "canopylabs/3b-de-ft-research_release"
ADAPTER_ID = "griffing52/orpheus-swiss-german-lora"

dtype = torch.float16 if torch.cuda.is_available() else torch.float32

# Load base model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=dtype)

if torch.cuda.is_available():
    base_model = base_model.to("cuda")

# Load fine-tuned adapter
model = PeftModel.from_pretrained(base_model, ADAPTER_ID)
model.eval()

Training Details

Training Data

The model was fine-tuned using paired Swiss German audio and transcripts structured in a SpeechT5-style metadata format.

Audio Format: 24kHz mono (WAV/FLAC/MP3) encoded into SNAC tokens.
Target Layout: Tokens packed into the native 7-token interleaved Orpheus speech architecture.

Training Procedure

Preprocessing

Transcripts and voice tags were converted into text-conditioning prompts bounded by specialized system tokens (SOH_TOKEN, EOT_TOKEN, EOH_TOKEN). Audio files were resampled to 24,000 Hz, passed through a hubertsiuzdak/snac_24khz codec model to extract 3 layers of acoustic features, and flattened into speech sequences.

Training Hyperparameters

Training regime: [More Information Needed] -Training regime: QLoRA 4-bit (nf4 quantization with double quantization enabled)
- Compute Precision: fp16 mixed precision
- LoRA Configuration:
  - r: 32
  - alpha: 64
  - dropout: 0.05
  - Target Modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- Optimization Strategy:
  - Learning Rate: 2e-4
  - LR Scheduler: Cosine
  - Warmup Ratio: 0.03
  - Epochs: 2
  - Per Device Train Batch Size: 1
  - Gradient Accumulation Steps: 24
  - Gradient Checkpointing: Enabled

Speeds, Sizes, Times

Total Training Steps: Concluded at checkpoint-216.

Evaluation

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

Hardware Type: GPU (Google Colab Environment)
Cloud Provider: Google Cloud

Technical Specifications [optional]

Framework versions

PEFT 0.18.1
Transformers 5.0.0+
PyTorch 2.10.0+cu128
SNAC 1.2.1