apptor-tiny-tts / README.md
apptor's picture
apptor-tiny-tts German v1 (step 82044, Vocos 60k)
b5c6ae8 verified
|
Raw History Blame Contribute Delete
4.43 kB
metadata
license: apache-2.0
language:
  - de
pipeline_tag: text-to-speech
library_name: onnx
tags:
  - tts
  - text-to-speech
  - german
  - onnx
  - cpu
  - tiny-voca
  - vocos

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.

An optional second stage, a Vocos vocoder fine-tuned on the two voices (13.5M parameters, 54 MB), re-synthesises the waveform and removes most of the buzzy sound of the tiny built-in generator. It is still real time on the CPU (~20x).

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-tiny-tts-vocos-de.pt optional: Vocos state dict, fine-tuned from charactr/vocos-mel-24khz MIT
tinyvoca/ German text front end (eSpeak NG phonemes, number normalisation), from tiny-voca Apache-2.0
speak.py minimal example Apache-2.0

Use

Needs the espeak-ng program (apt install espeak-ng; GPL-3.0, called as a separate process) and pip install onnxruntime numpy soundfile num2words.

python speak.py "Am 3. Mai um 18:30 Uhr kostet das Ticket 12,50 Euro." --voice M2 --out test.wav
# with the Vocos second stage (pip install torch vocos):
python speak.py "Hallo, wie geht es dir?" --voice F2 --vocos --out test.wav

ONNX inputs: tokens and stress (int64 [1, N], from tinyvoca.text.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: eSpeak NG de, 67 phoneme symbols (the English inventory of tiny-voca replaced by a German one), German number/date/time normalisation.
  • 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 optionalem Vocos-Vocoder. Gewichte und Code unter Apache 2.0, also auch kommerziell nutzbar, der Vocoder unter MIT. Keine offizielle tiny-voca-Version.

Credits