Add F5-TTS community model (M0 scaffolding; Habibi Arabic aliases) - #275

Merged

0xShug0 merged 28 commits into

0xShug0:mainfrom

tareko:community/f5-tts

Aug 22, 2026

Commits on Aug 18, 2026

Add F5-TTS community model scaffolding (M0)
Tarek Loubani
committed
19dee82
F5-TTS M2/M3: verified DiT parity + full inference pipeline
Tarek Loubani
committed
9776633
F5-TTS: CUDA inference for the DiT (12x speedup)
Tarek Loubani
committed
4db89f4

Commits on Aug 19, 2026

F5-TTS: graph reuse + GPU vocoder + CUDA graphs (0.75x RTF, real-time)
Tarek Loubani
committed
740c69f
F5-TTS: batched CFG single-pass + FP16 option + graph bucketing (0.31x RTF warm)
Tarek Loubani
committed
64f94ac
F5-TTS: framework-module graph + long-text chunking (review feedback)
Tarek Loubani
committed
4f4a119
F5-TTS: fix long-form pacing (chunk sizing by duration budget, ref-driven rate)
Tarek Loubani
committed
3b78bf9
F5-TTS: fix silent/garbled long-form output (three stacked bugs)
Tarek Loubani
committed
1de07be
F5-TTS: drop chunker test extraction artifact
Tarek Loubani
committed
5cb243e
F5-TTS: fix arena aliasing corruption (the actual garbled-output root cause)
Tarek Loubani
committed
295fbe9
F5-TTS: correct IRQ.wav transcript; drop chained references
Tarek Loubani
committed
41a7165
F5-TTS: checkpoint-varying vocab size + IRQ specialized model support
Tarek Loubani
committed
47abdc2
F5-TTS: fix text encoder to python semantics (padded + masked ConvNeXt)
Tarek Loubani
committed
796b883
F5-TTS: fix CFG null branch + corrupted e2e Arabic text; add parity/ASR tests
Tarek Loubani
committed
4f6e886

Commits on Aug 20, 2026

F5-TTS: wire session inference into audiocpp_server + docker-link fix
Tarek Loubani
committed
92d00a8
F5-TTS: fix periodic word dropping in long-form synthesis
Tarek Loubani
committed
38ee09d
F5-TTS: random seed per request when seed is unset (python parity)
Tarek Loubani
committed
1eca50d
F5-TTS: stop truncating reference audio without its transcript (ref leak)
Tarek Loubani
committed
ad1c580
F5-TTS: fix long mid-text pauses with pause-heavy references
Tarek Loubani
committed
4512ffb
F5-TTS: self-contained install from the repo (vocos package, habibi alias, spec fix)
Tarek Loubani
committed
90e70e2
voxcpm2: restore upstream 31-language list
Tarek Loubani
committed
cac0f8f
F5-TTS: revert ggml-alloc.c / ggml-cuda.cu to upstream (PR #275 feedback)
Tarek Loubani
committed
b3b9323

Commits on Aug 21, 2026

F5-TTS: GGUF packaging (PR #275 feedback)
Tarek Loubani
committed
0e44570
F5-TTS: fix bundled-vocos probe (slash namespaces) + point packages at trklou/audio.cpp
Tarek Loubani
committed
dbd4bb0

Commits on Aug 22, 2026

F5-TTS: tail pad for single-chunk synthesis (final-phoneme clipping)
Tarek Loubani
committed
184d6fd
F5-TTS: address PR #275 round-2 review (Windows CI, options, hygiene)
Tarek Loubani
committed
1a8bf22
F5-TTS: strip Arabic diacritics by default (harakat garbling)
Tarek Loubani
committed
1c55248
F5-TTS: widen single-chunk tail pad 20 -> 40 frames
Tarek Loubani
committed
249e3e9