Add Echo-TTS (community model) — draft, opening early per #54

[5uck1ess]

Draft — opening early per #54 to flag the area and get direction before writing the bulk of the port.

Adds Echo-TTS as a community model: an English
zero-shot voice-cloning TTS model. A 2.8B diffusion transformer generates 80-D latents in PCA space,
decoded to 44.1 kHz by the Fish S1-DAC autoencoder. Cloning needs a reference wav, no transcript.

Echo-TTS is on the candidate list in #34 (struck through, "contributions are welcome").

What's in this commit

Scaffolding only — no model math yet:

• model_specs/echo_tts.json, schema_version: 1, using the generic spec-backed loader (no loader.cpp)
• Session stub returning silence, so the family registers and links
• docs/community_models/echo_tts.md

Verified on top of current main: registered_loaders goes 44 → 45, and echo_tts: clon (offline) appears in --list-loaders.

Plan

Each milestone is gated on cosine ≥ 0.999 parity against the reference PyTorch implementation on a
fixed seed, plus a listening check.

Milestone	Scope
M0	Family registration + spec (this commit)
M1	GGUF conversion, DiT, PCA inverse, Fish decode
M2	Native speaker encoding (Fish encoder + RVQ) + long-form chunking
M3	Q8_0, RTF and VRAM evidence, docs

This PR stays in draft until the full evidence pack exists: build/run commands, parity numbers,
path-test results across safetensors/F16/Q8_0, RTF < 1.0, stable VRAM across repeated requests, and
generated wavs.

How long text is handled

Echo generates at most 640 latents (640 x 2048 / 44100 = 29.7215 s). That is a property of the
model, not the port: longer text is spoken faster to fit rather than truncated, and the tokenizer
hard-truncates past 768 UTF-8 bytes. Upstream's blockwise sampler subdivides that window rather
than extending it (sum(block_sizes) + continuation_length < 640), and upstream notes it "hasn't
been thoroughly tested".

So long text goes through the framework chunker, exactly as the rest of the repo does:

const int64_t text_chunk_size =
    engine::text::parse_text_chunk_size_override(request.options).value_or(kDefaultTextChunkSize);
const auto chunk_requests = runtime::chunk_text_request(request, text_chunk_size);
for (const auto & chunk_request : chunk_requests) { /* synthesize with cached conditioning */ }

engine::text::split_text_chunks is already consumed by 19 session.cpp files, and chatterbox
is the closest analogue — a clone family with cached speaker conditioning that concatenates chunk
audio via append_audio_buffer. Echo caches its speaker latent once per session, so timbre is
stable across seams by construction. Planned kDefaultTextChunkSize is 300 codepoints (~20 s
of English), leaving headroom before the model starts compressing; overridable at runtime.

I've deliberately not claimed long_form in capabilities — it appears nowhere in the C++, and
only 5 of 22 TTS/clone families declare it. The 17 that don't include chatterbox, fish_audio,
higgs_audio_tts, index_tts2, qwen3_tts, voxcpm2, and pocket_tts. Happy to add it if you'd
rather it be declared.

One question

Anything you'd want structured differently before there's a lot of code to restructure — file
layout, option naming, or whether this belongs in community_models at all.

Licence

Echo-TTS is CC-BY-NC-SA-4.0, and the restriction covers generated audio, not just the weights
(inherited from the Fish S1-DAC dependency). Flagging it explicitly rather than leaving it to be
inferred. There's precedent in-tree — fish_audio carries the identical restriction from the
identical dependency — and audio.cpp's own Apache 2.0 licence is unaffected, since weights are a
separate download. Documented in the model doc so users don't ship product on non-commercial output.

Why this model

Picked by comparing every model tracked in tts-bench — 62
local TTS models benchmarked across speed, objective scores, and blind human preference — against the
existing support table.

Measure	Echo-TTS	Field
Blind cloning Elo	1162	#3 of 40 (35 games, 738 cloning votes)
Speaker similarity	0.836	2nd of 41
UTMOS / WER	4.21 / 7.45 %	—
Frozen pairwise study	21-1-6	near-tied 1st of 28

Caveats worth stating: the cloning arena averages ~30 games per model, so gaps under ~100 Elo are
noise, and the ranking uses a single reference clip. Echo is top-3 on votes and 2nd on objective
SIM, which are independent measurements.

Compute profile suits a GGUF port — ~2.8B at 1.35× RTFx and 9.4 GB VRAM in PyTorch on an RTX 3090,
so there's real work to amortise.