from f5_tts.api import F5TTS

tts = F5TTS(
    model="F5TTS_v1_Base",
    lora_path="instavar/f5-tts-v1-lora-female01",
    device="cuda",
)

wav, sr, spec = tts.infer(
    ref_file="reference.wav",
    ref_text="",
    gen_text="Text to synthesize.",
    file_wave="output.wav",
)

Or via CLI:

f5-tts_infer-cli --lora_path instavar/f5-tts-v1-lora-female01 -r ref.wav -t "Text to synthesize."

Requires the instavar/f5-tts-lora-finetuning fork.

Training details

Parameter	Value
Base model	`F5TTS_v1_Base` (`model_1250000.safetensors`, SHA256 prefix `670900fd`)
Dataset	IMDA NSC FEMALE_01 (10,850 train samples, pinyin tokenizer)
LoRA rank	16
LoRA alpha	16
Target modules	`to_q`, `to_k`, `to_v`, `to_out.0`
Trainable params	2,883,584 / 339,980,388 (0.85%)
Learning rate	1e-4
Epochs	1 (~1,250 updates)
Hardware	NVIDIA RTX 3090 Ti (24 GB)
EMA	Disabled (standard for diffusion LoRA)

Adapter details

Framework: PEFT 0.18.1
Adapter size: 12 MB (vs ~2.7 GB for full fine-tuning checkpoint)
Merge strategy: merge_and_unload() at inference for zero overhead
F5-TTS version: 1.1.18 (instavar fork, commit dc1b18a)

License

MIT (same as upstream F5-TTS code). The base model pre-trained weights are CC-BY-NC due to the Emilia training dataset.