Parler-TTS Large v1

Parler-TTS is a lightweight text-to-speech model you control with natural-language descriptions like "a female voice with an Italian accent, speaking fairly quickly, very close recording". It was developed by researchers Dan Lyth and Simon King and released as a fully open-source package: datasets, preprocessing, training code, and weights are all public under a permissive license (per the model card).

Benchmarks (provider-reported, from the Parler-TTS paper — arxiv:2402.01912):

The paper evaluated the Parler-TTS approach (mini v0.1) against Meta's Audiobox and ground truth. Large-v1 was released later with no separate published eval table, so treat these as lower bounds for the family:

• Subjective listening test (30 native-English listeners, 5-point MOS, 95% CI): naturalness 3.92 ± 0.07 vs Audiobox 2.79 ± 0.09 vs ground truth 3.67 ± 0.09; description-relevance 3.88 ± 0.06 vs Audiobox 3.19 ± 0.06 vs ground truth 3.62 ± 0.06 (it beat ground truth because test-set descriptions contained label noise — the authors' own caveat).
• Objective (Torchaudio SQUIM, reference-less, 20 LibriTTS-R sentences): PESQ 3.84 ± 0.10 vs Audiobox 3.46 ± 0.16 (GT 4.15); STOI 0.996 vs 0.988; SI-SDR 26.53 dB vs 21.84 dB.
• No independent third-party reproduction of these numbers was found.

Practical notes: the training data is English-only (MLS_eng + LibriTTS-R), so non-English voices are out of scope. The key differentiator is the description interface — no reference-audio needed to pick a voice, unlike XTTS-style cloning.

Provider-reported (paper arxiv:2402.01912, mini v0.1 eval):

• MOS naturalness 3.92 ± 0.07 (Audiobox 2.79, ground truth 3.67) — 30 listeners, 5-point scale
• Description relevance 3.88 ± 0.06 (Audiobox 3.19, ground truth 3.62)
• PESQ 3.84 ± 0.10 (Audiobox 3.46); STOI 0.996; SI-SDR 26.53 dB

No independent benchmarks found. Large-v1 has no separately published eval.