reports/tts-head-to-head.md

Sovereign TTS head-to-head: the 1.7B Apache model beats the 4B research-license model on a 5090

Rig: one RTX 5090 32GB (sm_120) · both models bf16, neither compiled · Seed-TTS-eval EN (150 utterances) · voice-clone from each utterance's reference clip Models: Qwen3-TTS-12Hz-1.7B-Base (Apache-2.0, the qwen-tts package) vs fishaudio/s2-pro (4B dual-AR, Fish Audio Research License, from-source fish-speech) Metrics (4 axes): round-trip WER (synth → whisper-large-v3 → WER vs target), SIM-o (WavLM-large+ECAPA cosine, the Seed-TTS/F5-TTS metric), RTFx (audio seconds ÷ synth seconds), first-audio latency. Same harness, same subset, same judge for both.

The numbers (both n=150)

axis	Qwen3-TTS-1.7B	Fish-S2-Pro (4B)	Fish-S2-Pro compiled	winner
round-trip WER	0.6%	0.6%	0.7%	≈ tie
SIM-o (speaker clone)	0.699	0.625	0.621	Qwen
RTFx (× realtime)	2.22×	0.39×	1.59×	Qwen
first-audio latency	1.72s	9.74s	2.11s	Qwen

Both are equally intelligible. The small one is ~6× faster, clones a touch better, and answers ~6× sooner. On a consumer 5090, out-of-the-box, the 1.7B Apache model is the better sovereign pick.

The mechanism, not just the scores

• WER is a tie because both models are simply good. Round-trip WER lands at 0.6% for each — Fish's "sub-1% WER" claim holds, and Qwen matches it. Intelligibility is not where these two separate.
• Speed is where the architectures diverge — and where the deployment story lives. Fish-S2-Pro is a 4B dual-AR model whose serving stack is built for SGLang + torch.compile + datacenter cards (the vendor RTF<0.5 number is H100/H200-class). Run it the way you'd run any model out-of-the-box on a 5090 — bf16, no compile — and it generates at 0.39× realtime (≈3× slower than the audio it's making). Qwen3-TTS-1.7B, same conditions, runs at 2.22×. That's a 5.7× gap from model size + serving assumptions, not from quality.
• Latency compounds it. First audio at 1.72s (Qwen) vs 9.74s (Fish) — the 4B's longer generate makes it unusable for anything interactive without the compiled path.
• SIM-o: a real but modest edge to Qwen (0.699 vs 0.625). Both clone faithfully from a single reference; Qwen's embeddings sit closer to the prompt speaker.

Compiled-Fish: does the gap close? (the follow-up, run 2026-06-22)

Re-ran Fish through its api_server with --compile on the same 150 utterances, same sm_120 5090, everything else identical.

• Compile is a large, real lever on consumer hardware. RTFx goes 0.39× → 1.59× (4.1×) and first-audio 9.74s → 2.11s (4.6×). Fish moves from 3× slower than realtime to 1.6× faster. The vendor's "fast with torch.compile" direction is not H200-only; it lands on a 5090.
• Quality is untouched (WER 0.7%, SIM-o 0.621, both within noise of the out-of-box run) — confirming compile is a pure speed lever, not a quality trade.
• But it does not flip the verdict. Even compiled, Fish (1.59×, 2.11s) still loses both speed axes to the 4×-smaller Qwen3-TTS-1.7B (2.22×, 1.72s), which needs no compile step at all. The gap shrinks from 5.7× to 1.4×; it does not close.
• The cost is a one-time 110s compile warm-up at server startup before the first synth.

Caveats (read these before quoting the numbers)

• The main table is no-compile for both — the honest apples-to-apples out-of-the-box comparison. The compiled-Fish follow-up above now quantifies the --compile path: 4.1× faster (1.59×), still 1.4× behind the uncompiled 1.7B.
• Round-trip WER uses whisper-large-v3 as the judge — comparable between these two models on the same subset, not directly comparable to either vendor's own WER protocol.
• SIM-o is the standard WavLM-large+ECAPA (wavlm_large_finetune.pth), the same metric the Seed-TTS / F5-TTS / CosyVoice papers report.
• Licenses differ: Qwen3-TTS is Apache-2.0; Fish-S2-Pro is research/non-commercial. For a sovereign, ship-it stack that matters as much as the numbers.

Worth it if / not if

• Reach for Qwen3-TTS-1.7B if you want a fast, permissively-licensed voice-clone TTS that runs on one consumer GPU today. It is the out-of-the-box winner here on every axis that isn't a tie.
• Reach for Fish-S2-Pro if you want its fine-grained inline prosody control ([whisper], [excited], free-form tags) — features this intelligibility/speed bench doesn't measure — and you'll run the --compile path. Compiled, it is genuinely deployable (1.59× realtime, 2.11s first-audio), but it still trails the 1.7B on speed at 4× the size. Out-of-the-box without compile on a 5090, it is not the pick.

Repro

• Synthesis adapters: scripts/tts_synth_qwen3.py (qwen-tts, SDPA on sm_120), scripts/tts_synth_fish.py (fish-speech ormsgpack HTTP server). Scoring: scripts/tts_bench.py (soxr 16k resample, batched whisper round-trip, SIM-o). Metric layer: lib/tts/. Chart: scripts/chart_tts.py.
• sm_120 notes that cost time: Fish from-source needed pyaudio dropped (no PortAudio headers; mic-only, irrelevant to batch synth); torch 2.8.0+cu128 runs the 4B dual-AR + codec on sm_120; the SIM-o path needs a local s3prl cache with a wavlm-only hubconf + torchaudio-2.x shims (the legacy s3prl zoo breaks on set_audio_backend/sox_effects).