| F5-TTS-THAI v1 | Local GPU | Weights tagged CC-BY-4.0, but fine-tuned from F5-TTS base whose weights are CC-BY-NC (Emilia data): commercial use unclear | Free (our GPU) | Yes, from a ~4 s sample | 6–9 min per line (93–97×), load 8 s | Very accurate Thai. Voice = the repo's sample speaker (cloned). |
| KhanomTan v1.1 (YourTTS) | Local GPU or CPU | Apache-2.0 | Free | Yes (YourTTS speaker embedding), 4 built-in speakers | 1.3–2.6 s per line (0.36–0.46×), load 20 s | The only commercial-clean engine that runs faster than realtime on a CPU. Small model, voice quality below F5-class; misread line 3 (CER 0.47) and สมชาย → สมชัย. |
| Qwen3-TTS 1.7B CustomVoice | Local GPU | Apache-2.0 | Free (our GPU) | Yes (Base model), plus voice design | 30–53 s per line (6.8–8.3×), load 22 s | Thai is NOT in its 10 official languages (run with language=Auto), yet it is intelligible. English-speaker voice: listen for accent. Slower than realtime on an RTX 3070. |
| CosyVoice 2 0.5B (zero-shot) | Local GPU | Apache-2.0 | Free (our GPU) | Yes, zero-shot | Not measured (blocked on this machine) | Does NOT speak Thai: gibberish on every line (CER 0.79–1.34). Thai is not an official language. A community Thai checkpoint exists (isabeth/SE-Bridge-TTS, licence “other”, unverified): not tried. |
| Azure neural th-TH Niwat (via edge-tts) | Cloud | edge-tts uses Edge's free read-aloud endpoint: fine for this demo, NOT for production. Production = Azure AI Speech, same voices, paid | Azure per-character pricing | No (Azure custom voice is gated) | Generated in Microsoft's cloud, not on our CPU (~0.6–1.6 s per line incl. network) | Natural, fast. Shows the quality of Azure's Thai voices. |
| Azure neural th-TH Premwadee (via edge-tts) | Cloud | As above | As above | No | As above | Returned no audio for line 1 on every try (a male-polite “ครับ” line?); read การ์ด as กาด. |
| Meta MMS-TTS Thai | Local GPU or CPU | CC-BY-NC-4.0: NOT for commercial use | Free | No (one voice) | 0.7–1.0 s per line (0.14–0.20×), load 8 s | Very fast, robotic. Read การ์ด as กาด. Reference only. |
| ElevenLabs eleven_v3 (prototype voice) | Cloud | Commercial API | ~US$0.08 / 1k chars (prototype estimate) | Yes | Cloud | Not run on these lines: no key in aa-platform/.env. These are the prototype's own pre-made files (same voice and model, its own sentences). Line 3 is already respelled ก๊าด. |