FreyaTTS-small (Leyla, extended-coverage fine-tune)
A continued fine-tune of
freyavoice/freya-tts
(FreyaTTS-small, 183.2M params) on the
original Leyla voice, using voice-converted Common
Voice tr content to broaden text/phonetic coverage without changing the speaker identity.
This is an independent community fine-tune, not an official Freya release.
What this is
The base model is already voice-locked to Leyla, but its SFT corpus is a relatively narrow,
proprietary single-speaker set. This fine-tune keeps the Leyla identity and continues training on
a much larger, more diverse text corpus, by:
- Taking ~19,979 clips (~21.5h) from the highest-clip-count speaker in
Mozilla Common Voice tr 26.0 (CC0-1.0) — real, varied
Turkish sentences, not synthetic text.
- Converting every clip's timbre to the Leyla voice with seed-vc
(zero-shot voice conversion; diffusion-steps=90, length-adjust=1.0, inference-cfg-rate=0.8),
while keeping the original speaker's prosody/content and text transcript. Source audio is
silence/click/breath-trimmed adaptively (
librosa.effects.trim, top_db=40) before conversion.
- Full-parameter fine-tuning the original, untouched
freyavoice/freya-tts checkpoint on this
Leyla-timbre/diverse-content set — not on the raw Common Voice speaker's own voice.
Why this approach
A naive fine-tune on real Common Voice audio would overwrite Leyla's identity with the donor
speaker's voice (this is what full-parameter SFT does — see the base model's technical report,
section 3.5). Re-voicing the content into Leyla's timbre first lets the model absorb the text/
phonetic diversity of a large real-speech corpus while the identity signal in every training
example stays Leyla.
Checkpoint selection and identity check
Candidates were compared by synthesizing fixed sentences at multiple seeds and measuring F0
(pitch) mean/std, both against each other and against the unmodified base model as a
reference point:
| Checkpoint | F0 mean | F0 std |
|---|
freyavoice/freya-tts (base, reference) | 289.8 Hz | 19.30 Hz |
| this repo (step 3500 of 4000) | 294.7 Hz | 19.99 Hz |
| step 4000 (rejected) | 299.7 Hz | 30.05 Hz |
Step 3500 is within ~2% F0 mean and ~0.7 Hz F0 std of the untouched base model — the identity is
essentially preserved. Step 4000 had already started drifting (consistent with the base model's
own report finding that continued SFT eventually destabilizes the locked voice), so step 3500
was selected instead of the final checkpoint.
Architecture, inference code, and usage are unchanged from the base model — this repo only
replaces the weights.
Usage
Requires the
freyavoiceai/FreyaTTS inference code
(
pip install -r requirements.txt):
1from freyatts import FreyaTTS
2
3tts = FreyaTTS.from_pretrained("<this-repo-id>", device="cuda")
4wav = tts.synthesize("Merhaba, size nasıl yardımcı olabilirim?")
5tts.save_wav(wav, "output.wav")
Data provenance and license
Training text/content is derived from Common Voice tr 26.0, released by Mozilla under CC0-1.0
(public domain dedication). Only the linguistic content and prosody pattern of the donor
recordings is used; their timbre is discarded by the voice-conversion step, and every training
example's speaker identity is Leyla's, not the Common Voice contributor's. The donor speaker is
identified only by an anonymized client_id from the public dataset.
Model weights and code inherit the base model's Apache-2.0 license.
Limitations
Same as the base model: narrowband fidelity ceiling from the AudioVAE2's 16 kHz encode path,
digit-dense input requires spoken-form expansion at the text frontend, and long inputs can drift
without clause chunking. This fine-tune was not evaluated on the Freya-TR-Eval benchmark; no
WER/CER/MOS numbers are claimed here — the checkpoint selection above is an identity-consistency
check, not a quality/intelligibility benchmark.
Citation
