f5-tts-northern-english-ft

F5-TTS fine-tuned for Northern English (Bolton/Lancashire)

Fine-tuned from SWivid/F5-TTS (F5TTS_v1_Base/model_1250000.safetensors) on a small (~163 min) corpus of single-speaker British audio across three speakers from the Bolton / Lancashire region: Sara Cox, Maxine Peake, Diane Morgan.

What this is

A 9-epoch fine-tune of F5-TTS that produces strong Northern accent commitment with occasional phonetic failures. Specifically:

• ✅ Distinctly Northern (FOOT-STRUT collapse on common words: "London" → "Lundun"; BATH-rejection: "laughing" with short /æ/)
• ✅ Stronger accent commitment than the StyleTTS2 variant
• ✗ Occasional truncation of word endings ("sunshine" → "sunshinn")
• ✗ Occasional consonant deletion ("morning" → "monning")
• ✗ Occasional dropped function words

This is one of two checkpoints from the same project. See the StyleTTS2 variant for the cleaner / less-aggressive alternative.

Why epoch 9 specifically

We trained 10 epochs and listened to renders from epochs 4 through 10. Epoch 9 was the sweet spot: accent fully committed, but truncation / word-dropping less prominent than at epoch 10. Beyond epoch 10 we expect those phonetic failures to deepen.

The architectural reason F5 has this failure mode (and StyleTTS2 doesn't) is in the trade-off write-up: F5's single flow-matching loss has no explicit duration / alignment penalty, so late-epoch over-fitting can compress pace at the cost of phoneme completeness. StyleTTS2's eight constraints prevent this but at the cost of accent strength.

Usage

from f5_tts.api import F5TTS

# Load with the fine-tuned checkpoint
ft = F5TTS(model="F5TTS_v1_Base", ckpt_file="ckpt_epoch_9.pt")

# Render with any reference clip (we recommend a Sara Cox audiobook clip
# for Sara-flavoured output, or any Northern English reference for that
# speaker's voice)
audio, sr, _ = ft.infer(
    ref_file="ref.wav",
    ref_text="The exact transcript of ref.wav, including filler words.",
    gen_text="A tensor lives on the manifold.",
)

For the full minimal trainer pipeline that produced this checkpoint, see the companion gist.

Architecture and base

• Base: SWivid/F5-TTS — F5TTS_v1_Base — flow-matching CFM with 22-layer DiT (1024 dim, 16 heads).
• Fine-tune: 10 epochs at lr=5e-5 constant (no decay), batch_size=2, full corpus.

Training corpus composition

Important detail: the published F5TTS_v1_Base checkpoint is EMA-only

If you load the base checkpoint with a custom loader, strip the ema_model. prefix and use those weights as the start. See issue #1292 for the gory detail. This fine-tune was correctly initialised from the EMA weights.

Companion writeups

• Architecture trade-off (F5 vs StyleTTS2)
• How human feedback steers TTS fine-tuning
• Non-AVX2 CPU TTS compatibility notes
• Minimal F5-TTS trainer used for this checkpoint
• Project canonical site

Limitations

• The same training-corpus consent caveat as the StyleTTS2 model applies.
• This is the strong-accent variant; expect occasional phonetic imperfections. For cleaner output, use the StyleTTS2 variant.
• The checkpoint includes optimizer state (~2.7 GB) so the file is ~4 GB total. An inference-only stripped version is on the to-do list.

Citation

@misc{f5tts-northern-english-ft-2026,
  author = {netlinux-ai},
  title  = {F5-TTS fine-tuned for Northern English (Bolton/Lancashire)},
  year   = {2026},
  url    = {https://huggingface.co/grahamathf/f5-tts-northern-english-ft},
}