Vyvo/drifting-tts-en

Drifting TTS English (ARDrift)

An English text-to-speech model trained from scratch on 10,000 hours of synthetic speech (VoiceHub/echo-clones-10kh-en). It is the English version of ARDrift, the autoregressive model of kadirnar/drifting-tts.

• How it generates: a causal transformer produces the mel spectrogram 43 ms at a time. A small head turns noise into each piece in one pass. It is trained with the drifting objective (Deng et al., 2026), with Kyutai's learned temperature.
• Text: an English character frontend reads numbers, dates, money, clock times, abbreviations and acronyms itself. There is no phonemiser.
• Voices: the voices are those of the training data, chosen by speaker ID. There is no voice cloning.

Demo: Vyvo/drifting-tts-en-demo

Files

file	what	size
ardrift_en.pt	the acoustic model: 36 M text encoder, 121 M backbone, 20 M head	679 MB
bigvgan_en.pt	its vocoder: NVIDIA BigVGAN-v2 (24 kHz, 100 bands) fine-tuned on this model's own mels	450 MB
voices.json	eight voices picked from the 4,000 training speakers: a label and a speaker ID each	1 kB

The acoustic model's duration predictor was re-fitted to the mean duration of every token. With it, the model pauses at punctuation as long as the recordings do: 1.09 s per sentence against 1.19 s. The regressor it was trained with paused 0.45 s.

Usage

pip install "drifting-tts[bigvgan] @ git+https://github.com/kadirnar/drifting-tts"   # needs a CUDA build of PyTorch

import json

import soundfile as sf
from huggingface_hub import hf_hub_download

from drifting_tts.synthesize import Synthesizer

repo = "Vyvo/drifting-tts-en"
tts = Synthesizer(hf_hub_download(repo, "ardrift_en.pt"), "cuda", vocoder=hf_hub_download(repo, "bigvgan_en.pt"))
voices = json.load(open(hf_hub_download(repo, "voices.json")))["voices"]
wav, info = tts("Hello! It's 7:45, and the meeting moved to the 21st.", speaker=voices[0]["id"], temperature=0.5)
sf.write("hello.wav", wav.numpy(), 24000)

Use noise temperature 0.5 and no guidance (cfg_scale=1, the default).

Results

The 500 held-out sentences come from a source shard that is not among the 10,000 hours (val 200 + dev 300). The texts were never trained on; the voices were. Whisper large-v3 transcribes, UTMOSv2 predicts naturalness, and similarity is WavLM-ECAPA against the recording passed through the same vocoder.

	WER	CER	UTMOSv2	similarity
the recordings through this vocoder	1.57%	0.73%	3.294	–
this model	1.39%	0.61%	3.236	0.846
the same model with Vocos v2	1.42%	0.64%	2.930	–

The model is as intelligible as its synthetic training data, and its naturalness is close to the recordings' through the same vocoder. Details: docs/ENGLISH.md.

Limits

• Synthetic data: the model learns the voices and artefacts of the TTS model that generated its training data.
• The test set is from the same synthetic corpus: the texts were never seen, the voices were. There is no comparison with other English TTS systems on a common benchmark.
• Speakers by ID only: there is no voice cloning from a reference recording.
• Naturalness was measured only with automatic scores, not by listeners.

License

CC BY-NC-SA 4.0: non-commercial use only, and derivatives must share alike. The training speech was generated with EchoTTS (jordand/echo-tts-base), which is licensed CC-BY-NC-SA-4.0. The vocoder is fine-tuned from NVIDIA BigVGAN-v2 (MIT). The code is MIT (kadirnar/drifting-tts).