apptor-tiny-tts (German)
A tiny German text-to-speech model that runs on the CPU: 4.5M parameters, one 21 MB ONNX file, about 40x faster
than real time on four CPU threads. Two voices: F2 (female) and M2 (male), 24 kHz.
It is built on tiny-voca by n4ze3m: same architecture and training recipe, retrained from scratch for German. This is not an official tiny-voca release. Thanks to n4ze3m for the model and the idea.
A second stage, a Vocos vocoder fine-tuned on the two voices, re-synthesises the waveform and removes most of the buzzy sound of the tiny built-in generator. Since November 2026 it ships as an int8 ONNX file (17 MB), so no PyTorch is needed, and it is on by default.
New (2026-11-19): no eSpeak NG needed any more. A small letter-to-sound model (apptor-g2p-de.onnx, Apache-2.0)
replaces the eSpeak NG program, so the whole thing needs only onnxruntime, numpy and num2words, and it also
runs in the browser. There is also an Austrian-accented version with an accent slider:
apptor-tiny-tts-at, demo at
tiny-tts.apptor.at.
Listen: samples/demo.wav (with Vocos) and samples/demo-without-vocos.wav.
Files
| file | what | license |
|---|---|---|
apptor-tiny-tts-de.onnx |
the TTS model (text tokens -> 24 kHz audio) | Apache-2.0 |
apptor-tiny-tts-de.json |
sample rate, voices, phoneme symbols, limits | Apache-2.0 |
apptor-g2p-de.onnx / .json |
letters -> phonemes (conv + BiGRU + CTC, 12 MB), replaces eSpeak NG | Apache-2.0 |
apptor-vocos-de.int8.onnx |
sound filter: Vocos fine-tuned on the two voices, as ONNX (int8) | MIT |
apptor-tiny-tts-vocos-de.pt |
the same Vocos as PyTorch state dict (for vocos + torch users) |
MIT |
tinyvoca/ |
German text front end: number normalisation, tokens, G2P wrapper with a list of first names | Apache-2.0 |
speak.py |
Python example and small library | Apache-2.0 |
Use
pip install onnxruntime numpy num2words
python speak.py "Am 3. Mai um 18:30 Uhr kostet das Ticket 12,50 Euro." --voice M2 --out test.wav
python speak.py "Hallo, wie geht es dir?" --voice F2 --no-vocos --out test.wav # without the sound filter
from speak import Speaker
s = Speaker() # finds the models next to speak.py
audio = s.speak("Hallo!", voice="F2") # float32 numpy array, 24 kHz
ONNX inputs: tokens and stress (int64 [1, N], from tinyvoca.g2p_de.phonemize), voice (int64 [1], index into
voices), speed, pitch_shift, pitch_range (float32 [1]; 1.0 / 0.0 / 1.0 for normal). Outputs: audio
(float32 [1, samples]) and durations. At most 512 phoneme tokens per call; split longer text into sentences.
It is also a voice of flash-voice (services/tinytts_server.py,
same HTTP API as the other voices there), as an Apache-2.0 alternative to the German Phonon model, which is
CC BY-NC 4.0 (non-commercial only).
Training
- Data: about 12 hours of synthetic German speech, ~15,000 utterances, made with Qwen3-TTS 1.7B Base (Apache-2.0) with two of its built-in voices. Texts: Mozilla Common Voice German sentences (CC0). Utterances were screened with Whisper.
- Front end: trained on eSpeak NG
dephonemes, 67 phoneme symbols (the English inventory of tiny-voca replaced by a German one), German number/date/time normalisation. At run time the G2P model produces the same phonemes: it learned the eSpeak readings of ~480,000 words from the Common Voice sentences (98% of held-out words identical). - Model: tiny-voca recipe, 82k steps on one RTX 3090 (about 10 hours).
- Vocoder: Vocos fine-tuned for about 60k steps on an RTX 5060 Ti: input is tiny-voca's own teacher-forced waveform, target the recording.
- Data, training and tuning took about 12.5 hours on one Sunday.
Evaluation (Whisper large-v3-turbo, word error rate): 5.0% on 16 unseen sentences (F2 4.6%, M2 5.5%).
Limitations
- German only. English words are read with German pronunciation.
- The sound is clear but not studio quality; without Vocos it is noticeably buzzy.
- A word-final long "a" before a comma is sometimes drawn out ("Anna, ...").
- The voices are synthetic (Qwen3-TTS built-in voices), not clones of real people.
Deutsch
Ein winziges deutsches TTS-Modell für die CPU (4,5M Parameter, 21 MB ONNX), auf Basis von tiny-voca von n4ze3m nachtrainiert, mit Vocos-Klangfilter. Seit 19.11.2026 ohne eSpeak NG: nur onnxruntime, numpy und num2words, läuft auch im Browser. Österreichische Variante: apptor-tiny-tts-at. Gewichte und Code unter Apache 2.0, also auch kommerziell nutzbar, der Vocoder unter MIT. Keine offizielle tiny-voca-Version.
Credits
- tiny-voca by n4ze3m: architecture, training recipe, text front end (Apache-2.0)
- Vocos,
charactr/vocos-mel-24khz(MIT) - Qwen3-TTS (Apache-2.0): training speech
- Mozilla Common Voice German (CC0): texts
- eSpeak NG (GPL-3.0): phonemes for training and for teaching the G2P model; not included and not needed at run time. num2words (LGPL-2.1, installed separately)