Nabra-82M for sherpa-onnx
Arabic (MSA) TTS models packaged for sherpa-onnx, derived from Nabra-82M-v0.1.
Files
| File | Precision | Size |
|---|---|---|
model.int4.onnx |
INT4-QAT + notch FIR baked in | 57 MB |
model.int8.onnx |
INT8 weight-only + notch FIR | 83 MB |
model.fp16.onnx |
FP16 storage + notch FIR | 163 MB |
tokens.txt |
phoneme vocabulary (116 symbols) | 1 KB |
voices.bin |
style vectors [510, 1, 256] |
512 KB |
The iSTFT image-tone notch (4.8/9.6 kHz) is baked into every model as a final Conv1d, so no player-side post-processing is needed.
Usage (Python)
import sherpa_onnx
tts = sherpa_onnx.OfflineTts(
sherpa_onnx.OfflineTtsConfig(
model=sherpa_onnx.OfflineTtsModelConfig(
kokoro=sherpa_onnx.OfflineTtsKokoroModelConfig(
model="model.int4.onnx",
voices="voices.bin",
tokens="tokens.txt",
data_dir="espeak-ng-data",
),
),
)
)
audio = tts.generate("السلام عليكم", sid=0, speed=1.0)
Also works with sherpa-onnx command-line tools, Android/iOS APIs, Kotlin, Swift, Java, C#, Go, Dart, Rust and WebAssembly — anything sherpa-onnx supports.
Notes
- Single voice (
af_msa, speaker id 0), Modern Standard Arabic, 24 kHz output - Sequence cap: 510 tokens; espeak-ng
arvoice is used for G2P (bundledespeak-ng-datafrom your sherpa-onnx install or the official releases) - INT4 has a faint noise floor under speech; use INT8 for strictly clean audio
- The graph's SineGen noise node is seeded (42), but draws differ across CPU/GPU runtimes
License
Apache-2.0, following the base model.