SupraTTS-0.1-Beta is an English single-speaker text-to-speech model from SupraLabs, released under the MIT license as the successor to LH-Tech-AI's Flare-TTS-v1.5. It is a Glow-TTS acoustic model trained from scratch on LJSpeech-1.1 (a single English female speaker, about 24 hours) in Coqui-TTS, paired with a HiFi-GAN v1 vocoder. The acoustic model has about 29.6M parameters and the vocoder about 14M; the config lists 12 flow blocks, 192 hidden channels, a 6-layer relative-position transformer encoder, 22050 Hz audio and 80 mel bands. Compared with v1.5 the dataset, encoder and decoder are unchanged. The changes are a fix for a learning-rate scheduler bug (it stepped per epoch instead of per step, so v1.5 never reached its target learning rate), blank tokens between input tokens, a stochastic duration predictor borrowed from VITS in place of the deterministic one, bf16 training of the acoustic model with a Triton kernel for Monotonic Alignment Search, and a HiFi-GAN vocoder trained longer on ground-truth mels (about 103,000 steps). The developer evaluated it on 10 out-of-domain sentences plus 10 held-out LJSpeech sentences with 5 seeds each, using UTMOS as an automatic MOS estimate and Whisper word error rate as a mispronunciation check. Real recordings score UTMOS 4.34, F0-Std 3.51 semitones and WER 2.58%; Flare-TTS-v1.5 scores 3.05, 1.87 and 3.72% with a speech rate of 1.087x the recordings; SupraTTS scores 3.59, 1.01 and 3.64% at 1.004x. This is a self-evaluation on a small sentence set. The developer states that pitch is stable but flatter than a real speaker, with F0-Std well below the recordings, and that fine-tuning HiFi-GAN on the acoustic model's own mels (GTA) lowered UTMOS to 3.34 with noisier pitch, so the release uses the ground-truth-mel vocoder only. To run it, download infer_v2.py, flare_glowtts.py, the two config files, model.pth and vocoder.pth, install Coqui-TTS with PyTorch and CUDA on Python 3.10 or newer, edit the TEXT variable and run infer_v2.py, which writes output_v2.wav. It has one voice, English only, and no voice cloning; multi-speaker support and cloning appear only on the developer's roadmap for a later version.