nibor1896/F5-TTS-German-ONNX

F5-TTS-German — ONNX

Pure-ONNX export of the German F5-TTS voice-cloning model hvoss-techfak/F5-TTS-German, runnable with plain ONNX Runtime — no PyTorch, no Python required at inference time. This makes it easy to embed German text-to-speech in native applications (C#, C++, …) with only the ONNX Runtime dependency.

Exported with DakeQQ/F5-TTS-ONNX.

Why this export exists

hvoss-techfak/F5-TTS-German is a fine-tune of the F5TTS_Base (v0) architecture (text_mask_padding=False, pe_attn_head=1). DakeQQ's exporter was hard-wired for F5TTS_v1_Base, so exporting a v0 checkpoint produced fluent-but-wrong "word-salad" in the target language.

The root cause and fix were contributed upstream in DakeQQ/F5-TTS-ONNX#74 (merged):

• AttnProcessor now honors pe_attn_head — for v0 checkpoints RoPE is applied to the first head only, not all heads.
• TextEmbedding now honors mask_padding — v0 checkpoints must not zero out padding positions.

These files were exported from the fixed exporter with --omegacfg_path .../F5TTS_Base.yaml.

Files

File	Size	Role
F5_Preprocess.onnx	~65 MB	Reference audio (int16) + text ids → mel/text conditioning, RoPE tables, initial noise
F5_Transformer.onnx	~1.26 GB	One denoising step of the DiT (run in an NFE loop, default 32 steps)
F5_Decode.onnx	~60 MB	Denoised mel → 24 kHz int16 waveform (vocoder + ISTFT)
vocab.txt	—	Character/token vocabulary (unchanged from the source checkpoint)

Inference (pipeline order)

1. Tokenize the reference text + target text to text_ids (int32). For German, a plain character → vocab-index mapping is sufficient (jieba/pinyin segmentation is not required).
2. Load the reference voice clip as 24 kHz mono int16.
3. F5_Preprocess → noise, rope_cos_q, rope_sin_q, rope_cos_k, rope_sin_k, cat_mel_text, cat_mel_text_drop, ref_signal_len.
4. Loop F5_Transformer for N steps (default 32), feeding the step index into time_step.1 and the running noise/denoised back in.
5. F5_Decode(denoised, ref_signal_len) → int16 audio @ 24 kHz.

A reference implementation is F5-TTS-ONNX-Inference.py in the DakeQQ/F5-TTS-ONNX repo.

Verification

The exported models were verified end-to-end (pure ONNX Runtime) on an umlaut-heavy German sentence and transcribed with faster-whisper large-v3: language de, confidence 1.00, exact target transcript.

License & attribution

• Model weights derive from hvoss-techfak/F5-TTS-German and inherit its CC-BY-NC-4.0 (non-commercial) license.
• Base architecture: F5-TTS (SWivid et al.).
• ONNX export tooling: DakeQQ/F5-TTS-ONNX.

If you use this, please credit the original hvoss-techfak/F5-TTS-German checkpoint.