piper-de-de-thorsten-medium

Medium-size German male voice by Thorsten Mueller.

Model Details

Note: Piper uses the terms "medium", "high", etc. to refer to model size, not output quality. Medium models (~63 MB, ~15M params) and high models (~114 MB, ~28M params) both produce 22.05 kHz audio.

Usage

With piper-tts (GPL)

from piper import PiperVoice

voice = PiperVoice.load("model.onnx")
for chunk in voice.synthesize("Hello, this is a test."):
    # chunk.audio_float_array contains float32 audio
    pass

Standalone ONNX (MIT — no piper-tts dependency)

Requires espeak-ng installed (brew install espeak-ng / apt install espeak-ng).

import json, subprocess, numpy as np, onnxruntime as ort, soundfile as sf
from huggingface_hub import hf_hub_download

model_id = "Trelis/piper-de-de-thorsten-medium"
onnx_path = hf_hub_download(model_id, "model.onnx")
config_path = hf_hub_download(model_id, "model.onnx.json")

with open(config_path) as f:
    config = json.load(f)

session = ort.InferenceSession(onnx_path, providers=["CPUExecutionProvider"])
phoneme_id_map = config["phoneme_id_map"]
espeak_voice = config["espeak"]["voice"]

def phonemize(text, voice):
    out = subprocess.run(
        ["espeak-ng", "-v", voice, "-q", "--ipa=2", "-x", text],
        capture_output=True, text=True,
    ).stdout.strip()
    return [list(line.replace("_", " ")) for line in out.split("\n") if line.strip()]

def to_ids(phonemes, pmap):
    ids = [pmap["^"][0], pmap["_"][0]]
    for p in phonemes:
        if p in pmap:
            ids.extend(pmap[p])
            ids.append(pmap["_"][0])
    ids.append(pmap["$"][0])
    return ids

text = "Hello, this is a test."
audio_chunks = []
for sentence in phonemize(text, espeak_voice):
    ids = to_ids(sentence, phoneme_id_map)
    if len(ids) < 3:
        continue
    audio = session.run(None, {
        "input": np.array([ids], dtype=np.int64),
        "input_lengths": np.array([len(ids)], dtype=np.int64),
        "scales": np.array([
            config["inference"]["noise_scale"],
            config["inference"]["length_scale"],
            config["inference"]["noise_w"],
        ], dtype=np.float32),
    })[0]
    audio_chunks.append(audio.squeeze())

audio = np.concatenate(audio_chunks).astype(np.float32)
sf.write("output.wav", audio, config["audio"]["sample_rate"])

Fine-tuning

You can fine-tune this model on your own voice data using Trelis Studio. Piper models can be trained on custom datasets to create personalized voices.

Attribution

Trained on Thorsten Voice dataset by Thorsten Mueller. Fine-tuned from lessac medium.

Re-hosted from rhasspy/piper-voices. Original voice: de_DE-thorsten-medium