Kartoffel Orpheus 3B German Synthetic V0.1

🚀 Kartoffel-3B (Based on Orpheus-3B) - Synthetic

This is a German text-to-speech (TTS) model family based on Orpheus-3B. It offers high - quality speech synthesis with support for multiple speakers and various emotional expressions.

🚀 Quick Start

The following steps and code example show you how to use the Kartoffel - 3B synthetic model for text - to - speech synthesis.

✨ Features

• Multiple Speakers: The model can generate speech using various speaker identities from predefined speakers.
• Varied Expressions: Capable of generating speech with different emotional tones and expressions based on the input text.

📦 Installation

No specific installation steps are provided in the original document.

💻 Usage Examples

Basic Usage

import torch
import torchaudio.transforms as T
import os
import torch
from snac import SNAC

from peft import PeftModel
import soundfile as sf
from transformers import AutoModelForCausalLM, AutoTokenizer


model = AutoModelForCausalLM.from_pretrained(
    "SebastianBodza/Kartoffel_Orpheus-3B_german_synthetic-v0.1",
    device_map="auto",
)

tokenizer = AutoTokenizer.from_pretrained(
    "SebastianBodza/Kartoffel_Orpheus-3B_german_synthetic-v0.1",
)

snac_model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
snac_model = snac_model.to("cuda")

chosen_voice = "Martin"

prompts = [
    'Tief im verwunschenen Wald, wo die Bäume uralte Geheimnisse flüsterten, lebte ein kleiner Gnom namens Fips, der die Sprache der Tiere verstand.',
]

def process_single_prompt(prompt, chosen_voice):
    if chosen_voice == "in_prompt" or chosen_voice == "":
        full_prompt = prompt
    else:
        full_prompt = f"{chosen_voice}: {prompt}"
    start_token = torch.tensor([[128259]], dtype=torch.int64)
    end_tokens = torch.tensor([[128009, 128260]], dtype=torch.int64)

    input_ids = tokenizer(full_prompt, return_tensors="pt").input_ids
    modified_input_ids = torch.cat([start_token, input_ids, end_tokens], dim=1)

    input_ids = modified_input_ids.to("cuda")
    attention_mask = torch.ones_like(input_ids)

    generated_ids = model.generate(
        input_ids=input_ids,
        attention_mask=attention_mask,
        max_new_tokens=4000,
        do_sample=True,
        temperature=0.6,
        top_p=0.95,
        repetition_penalty=1.1,
        num_return_sequences=1,
        eos_token_id=128258,
        use_cache=True,
    )

    token_to_find = 128257
    token_to_remove = 128258

    token_indices = (generated_ids == token_to_find).nonzero(as_tuple=True)

    if len(token_indices[1]) > 0:
        last_occurrence_idx = token_indices[1][-1].item()
        cropped_tensor = generated_ids[:, last_occurrence_idx + 1 :]
    else:
        cropped_tensor = generated_ids

    masked_row = cropped_tensor[0][cropped_tensor[0] != token_to_remove]
    row_length = masked_row.size(0)
    new_length = (row_length // 7) * 7
    trimmed_row = masked_row[:new_length]
    code_list = [t - 128266 for t in trimmed_row]

    return code_list


def redistribute_codes(code_list):
    layer_1 = []
    layer_2 = []
    layer_3 = []
    for i in range((len(code_list) + 1) // 7):
        layer_1.append(code_list[7 * i])
        layer_2.append(code_list[7 * i + 1] - 4096)
        layer_3.append(code_list[7 * i + 2] - (2 * 4096))
        layer_3.append(code_list[7 * i + 3] - (3 * 4096))
        layer_2.append(code_list[7 * i + 4] - (4 * 4096))
        layer_3.append(code_list[7 * i + 5] - (5 * 4096))
        layer_3.append(code_list[7 * i + 6] - (6 * 4096))

    codes = [
        torch.tensor(layer_1).unsqueeze(0),
        torch.tensor(layer_2).unsqueeze(0),
        torch.tensor(layer_3).unsqueeze(0),
    ]
    codes = [c.to("cuda") for c in codes]

    audio_hat = snac_model.decode(codes)
    return audio_hat


for i, prompt in enumerate(prompts):
    print(f"Processing prompt {i + 1}/{len(prompts)}")
    with torch.no_grad():
        code_list = process_single_prompt(prompt, chosen_voice)
        samples = redistribute_codes(code_list)


    audio_numpy = samples.detach().squeeze().to("cpu").numpy()
    sf.write(f"output_{i}.wav", audio_numpy, 24000)
    print(f"Saved output_{i}.wav")

Advanced Usage

The basic usage code already covers most of the key steps. For more advanced usage, you can adjust the parameters in the model.generate function, such as max_new_tokens, temperature, top_p, etc., to get different speech generation results.

📚 Documentation

Model Overview

This is a German text - to - speech (TTS) model family based on Orpheus - 3B.

Two main versions are available:

1. Kartoffel - 3B - Natural: Fine - tuned primarily on natural human speech recordings, aiming for realistic voices. The dataset is based on high - quality German audio, including permissive podcasts, lectures, and other OER data that were processed with an Emilia styled pipeline.
2. Kartoffel - 3B - Synthetic: Fine - tuned using synthetic speech data, with emotions and different outbursts. The dataset consists of a diverse set of emotions with 4 different speakers.

This is currently the synthetic version for synthetic sounding speakers, but with added emotion and outburst support.

Available Speakers & Expressions for the synthetic Version

Speakers

• Martin
• Luca
• Anne
• Emma

Emotions

To add emotions the following ones are used:

• Neutral
• Happy
• Sad
• Excited
• Surprised
• Humorous
• Angry
• Calm
• Disgust
• Fear
• Proud
• Romantic

To use them add them behind the speaker name like [Speaker_name] - [Emotion]: [German text] for example for the speaker Martin and the Emotion sad, the correct template would be:

Martin - Sad: Oh ich bin sooo traurig. 

Outbursts

The following outbursts are working:

• haha
• ughh
• wow
• wuhuuu
• ohhh

You can either directly use them in the text or place them in tags. Keep in mind to use the exact text from the variations.

🔧 Technical Details

No specific technical details (more than 50 words of detailed technical description) are provided in the original document.

📄 License

The license of this model is llama3.2.