Kokoro 82M German (Victoria) — ONNX

kikiri-tts/kikiri-german-victoria — the German female voice Victoria Asztaller, a StyleTTS2 stage-2 fine-tune of Kokoro-82M by kikiri-tts — converted to the same end-to-end ONNX graphs as soniqo/Kokoro-82M-ONNX for on-device German TTS on Android.

German is not a voice of the multilingual Kokoro bundle: each German voice is a separate weight set trained on German phonemes only, so this bundle carries exactly one voice. The other voice, Martin, is published as soniqo/Kokoro-82M-German-Martin-ONNX.

Model

Property	Value
Parameters	82M
Architecture	Non-autoregressive, single-pass (BERT → duration → alignment → prosody → iSTFTNet)
Format	ONNX (opset 18)
Weights	162 MB (FP16 storage → FP32 compute, shared external data)
Sample rate	24 kHz
Max phonemes	128
Max audio	5.0 s full graph, 3.0 s short-turn graph
Voice	`df_victoria` (female)
Language	German (de)

Graph variants

Graph	Audio tensor	Intended use
`kokoro-e2e.onnx`	`[1, 1, 120000]` (5.0 s)	Full-capacity synthesis
`kokoro-e2e-realtime.onnx`	`[1, 1, 72000]` (3.0 s)	Default short voice-agent replies

Both graphs reference the same kokoro-e2e.onnx.data, so switching profiles is a small graph download.

Files

File	Description
`kokoro-e2e.onnx`	Full-capacity model graph
`kokoro-e2e-realtime.onnx`	3.0 s short-turn graph; shares the weights file below
`kokoro-e2e.onnx.data`	Model weights (external data, FP16 storage)
`vocab_index.json`	The upstream hexgrad IPA → token table (space = 16, ʦ = 20)
`dict_de.json`	German pronunciation dictionary (49,578 entries) in the voice's phoneme convention
`voices/df_victoria.bin`	Voice embedding (256 × float32)

Phonemes

The voice was trained on espeak-ng 1.52 German phonemes as post-processed by misaki (the kikiri-tts training front-end): I/W for ai/au, ɔø for eu, ʦ/ʧ/ʤ affricates, ɜ for a vocalic -er, the short ü written y, stress marks before the stressed vowel. dict_de.json holds the 50,000 most frequent German words in exactly that convention plus the kikiri brand/tech overrides. speech-core and speech-swift reproduce the rest of the front-end without espeak: number, date, time, currency, unit and abbreviation normalization, inflected and compound forms of dictionary words, and rule-based G2P.

This bundle's vocab_index.json holds hexgrad's exact token table, the same one the multilingual bundle has shipped since it was corrected on 2026-10-22.

Tensors

Input

Name	Shape	Type	Description
`input_ids`	`[1, 128]`	int64	Phoneme token IDs (zero-padded)
`attention_mask`	`[1, 128]`	int64	1 for real tokens, 0 for padding
`ref_s`	`[1, 256]`	float32	Voice style embedding
`speed`	`[1]`	float32	Speed factor (1.0 = normal)
`random_phases`	`[1, 9]`	float32	Initial harmonic phases (uniform [0,1))

Output

Name	Shape	Type	Description
`audio`	`[1, 1, 120000]` (`[1, 1, 72000]` realtime)	float32	Raw PCM waveform (24 kHz)
`audio_length_samples`	`[1]`	int64	Valid sample count (trim audio to this)
`pred_dur`	`[1, 128]`	float32	Predicted phoneme durations

Usage

import json
import numpy as np
import onnxruntime as ort

sess = ort.InferenceSession("kokoro-e2e-realtime.onnx")
vocab = json.load(open("vocab_index.json"))["vocab"]
phonemes = "ɡˈuːtən tˈɑːk, viː ɡˈeːt ɛs ˈiːnən?"  # from dict_de.json / speech-core
ids = [0] + [vocab[c] for c in phonemes if c in vocab] + [0]

input_ids = np.zeros((1, 128), dtype=np.int64); input_ids[0, :len(ids)] = ids
attention_mask = np.zeros((1, 128), dtype=np.int64); attention_mask[0, :len(ids)] = 1
voice = np.fromfile("voices/df_victoria.bin", dtype=np.float32).reshape(1, 256)

audio, length, _ = sess.run(None, {
    "input_ids": input_ids,
    "attention_mask": attention_mask,
    "ref_s": voice,
    "speed": np.ones(1, dtype=np.float32),
    "random_phases": np.random.rand(1, 9).astype(np.float32),
})
pcm = audio.flatten()[:int(length[0])]

On Android: ModelManager.ensureTtsModels(context, kokoroPack = KokoroPack.GERMAN_VICTORIA) then synthesize(text, "de").

Source

Converted from kikiri-tts/kikiri-german-victoria (revision ce81e200ff92, Apache 2.0) with the Soniqo end-to-end Kokoro export (FluidInference-aligned SineGen, masked AdaIN) and FP16 weight storage. Voice embedding: row 128 of the published [510, 1, 256] voicepack.

soniqo/Kokoro-82M-German-Victoria-ONNX