tedash34/ema-lightning-onnx

EMA Lightning (ONNX + WebAssembly normaliser)

EMA Lightning, a Turkish text-to-speech model with 8.6M parameters and 48 kHz output, exported to ONNX so it can run on ONNX Runtime, including in the browser with ONNX Runtime Web. The repository also carries normalizer-tr compiled to WebAssembly, the text normaliser that reads numbers, dates, money, units and abbreviations aloud.

These are the files delicious-cipetpet packs into ema-lightning-fp32.cpet. The weights are the original model's, unchanged. Nothing was retrained or quantised.

Files

File	What it is	SHA-256
text.onnx	letters → text features h and durations dur	bf6bc2bb7bcd7f61c2d1e58fced4e58b92ea6806e07befcabf0824d7d33a8ab8
sound.onnx	features + frame timeline + noise → latents (4 flow steps inside the graph)	3924d67f00c65d6b4c0632cd3c152fef9987df4e29f738533d357036ce3f6420
decoder.onnx	latents → 48 kHz audio (1920 samples per frame)	8246d10b495a5ead733f5478b7c732dc39f4662e08b314cb84a4758bbc1db13d
config.json	vocabulary and constants (latent_dim, steps, hop, sample_rate, fps)	293a43a887b7cbdecd458e3961bffb440bb385957eaec7c14a35378fa6f1de8b
normalizer_tr_wasm_bg.wasm	normalizer-tr 0.8.0 (fallback policy) via wasm-bindgen 0.2.129	946ba22f3f610357d192f84963d9b446a3828c25b847f29c03bca0cf64f9a341
ema-lightning.json	where every file comes from, in machine-readable form

SHA256SUMS lists the same hashes.

Graph signatures (fp32, opset 17, dynamic letters and frames axes):

text.onnx     ids int64 [1,L]                                           -> h f32 [1,L,C], dur f32 [1,L]
sound.onnx    h, dur, cw int64 [1,L], wstart int64 [1,L],
              fw int64 [1,T], fp f32 [1,T], noise f32 [1,4,T,64]        -> latents f32 [1,T,64]
decoder.onnx  z f32 [1,N,64]                                            -> audio f32 [1,N*1920]

Where the files come from

• Weights: ema.pt and decoder.pt of canberkkkkkk/ema-lightning at commit 7a6ba1ad216bb2f1da9863f80ac8770a6a807632.
• ONNX: exported with web/export_onnx.py from the ema-lightning browser port. The script checks every stage against PyTorch, and the export agrees with it to at least 87 dB (audio: 119 dB). The export is reproducible: running it again with Python 3.12, torch 2.14.1 and onnx 1.23.2 gives these exact bytes.
• Normaliser: normalizer-tr 0.8.0 with the fallback policy, compiled for wasm32-unknown-unknown with Rust 1.94.0 and wasm-bindgen 0.2.129. ema-lightning 1.1.2 accepts normalizer-tr 0.6 to 0.x, so text is read the way that release reads it. For example, iPhone'umdan becomes "ayfonumdan" and SON DAKİKA is read as words.

The model code that 1.1.2 changed since the export (loading checkpoints with weights_only=True, how the CPU scheduler batches windows) does not affect the graphs.

Using them

With delicious-cipetpet in the browser or in Node:

import { EmaLightning } from "cipetpet";

const tts = await EmaLightning.load("/packs/ema-lightning-fp32.cpet", { wasmPaths: "/ort/" });
const { audio, sampleRate } = await tts.synthesize("Merhaba, size nasıl yardımcı olabilirim?");

To run the graphs yourself, follow ema-lightning's engine.py. The text is normalised, then cut into pieces of at most about 10 seconds. Each piece goes through text.onnx. The durations become a frame timeline: frame counts per word, rounded half to even, and each frame's position inside its word. sound.onnx turns that timeline into latents, and decoder.onnx decodes them in windows with 8 frames of context on each side.

Limitations

These are the original model's: one voice, Turkish only, naturalness around UTMOS 3.3. The browser port seeds its noise with its own generator (mulberry32 + Box-Muller), not torch's, so a seed gives different audio here than in the Python package.

Credits and licence

• EMA Lightning by Canberk Aslan, Apache-2.0: model, code.
• normalizer-tr by Erdem Tuna, Apache-2.0: github.com/erdemtuna/normalizer-tr.

This export is distributed under the same licence, Apache-2.0 (see LICENSE).