rlx-tiny-tts

TinyTTS (tronghieuit/tiny-tts) — a MeloTTS / VITS2-style English text-to-speech model — running natively on RLX at 44.1 kHz. The four exported ONNX subgraphs (text_encoder, duration_predictor, flow, decoder) are imported into rlx-ir HIR and compiled per device; the monotonic-alignment + latent-sampling glue is reimplemented in Rust.

Quick start

cargo run -p rlx-tiny-tts --release -- \
  --data weights/tiny-tts-rlx --text "The weather is nice today." --out out.wav
# [--device cpu|metal|mlx|cuda|rocm|gpu] [--speaker MALE] [--speed 1.0] [--seed 1234]

--data is an RLX TinyTTS bundle (see scripts/export_tiny_tts.py): config.json, onnx/{text_encoder,duration_predictor,flow,decoder}.onnx, and a frontend/ asset dir. With no --device, the bin picks the best available accelerator (Metal → MLX → wgpu, else CPU).

Public API

use rlx_tiny_tts::{TinyTts, InferOpts, Device};

let model = TinyTts::load_from_dir(std::path::Path::new("weights/tiny-tts-rlx"))?;
let opts = InferOpts::from_config(model.config());

// Full pipeline: raw text → 44.1 kHz mono waveform, every graph on `device`.
let wav = model.synthesize_on("Hello, world!", Device::Metal, &opts)?;
// or `model.synthesize(text, &opts)` for the CPU backend.
println!("{} samples @ {} Hz", wav.samples.len(), wav.sample_rate);

Wav { samples, sample_rate } is the output; text_to_ids exposes the raw (phone, tone, lang) ids. The English text frontend (CMUdict + g2p_en + tagger + BERT) is reused byte-identically from rlx-inflect-nano (re-exported as rlx_tiny_tts::frontend).

Backends

The ONNX graphs compile per device, so TinyTTS runs on every RLX backend:

Convenience bundles: apple-silicon, nvidia-gpu, amd-gpu, all-backends.

Examples

keystone, flow_probe, tenc_probe, dec_probe, compile_all, debug_shapes under examples/ exercise individual subgraphs (single-stage parity / shape debugging).