ampixa/ratna-asr

Ratna

Ratna is a streaming Nepali speech recognition model from Ampixa Labs. It has 31,963,576 parameters and ships as a 35.10 MB int8 ONNX bundle that runs on one CPU thread or in a browser through WebAssembly. Try it at ampixa.com/ratna.

Files

File	Bytes	Purpose
encoder.int8.onnx	32,667,500	Causal Zipformer2 encoder
decoder.int8.onnx	1,333,266	Stateless predictor
joiner.int8.onnx	1,058,185	Joiner
tokens.txt	38,953	2,048 BPE tokens
web/model-0.data, web/model-1.data	18,000,000 and 17,097,904	The same model packed for the sherpa-onnx WebAssembly runtime

Use it

import sherpa_onnx
import soundfile as sf
from huggingface_hub import snapshot_download

path = snapshot_download("ampixa/ratna-asr", allow_patterns=["*.onnx", "tokens.txt"])
recognizer = sherpa_onnx.OnlineRecognizer.from_transducer(
    tokens=f"{path}/tokens.txt",
    encoder=f"{path}/encoder.int8.onnx",
    decoder=f"{path}/decoder.int8.onnx",
    joiner=f"{path}/joiner.int8.onnx",
    num_threads=1,
    sample_rate=16000,
    feature_dim=80,
    decoding_method="greedy_search",
    model_type="zipformer2",
)

audio, sr = sf.read("clip.wav", dtype="float32")  # mono, 16 kHz
stream = recognizer.create_stream()
stream.accept_waveform(sr, audio)
stream.accept_waveform(sr, [0.0] * int(0.5 * sr))
stream.input_finished()
while recognizer.is_ready(stream):
    recognizer.decode_stream(stream)
print(recognizer.get_result(stream))

How it works

16 kHz audio becomes an 80-dimensional log-mel filterbank every 10 ms. A Conv2d layer reduces it to 50 frames per second. A causal Zipformer2 encoder with six stacks runs at 50, 25, 12.5, 6.25, 12.5 and 25 Hz. A stateless predictor and joiner emit BPE tokens with greedy search. The model reads 320 ms chunks with 2.56 s of left context and no look-ahead.

Results

Test	CER %	WER %
PV read speech, 100 sentences	16.33	39.26
Fresh read speech	27.15	48.91
FLEURS	28.03	55.94

The real-time factor is 0.0273 on one thread of an Apple M2 Ultra, with a peak of 205 MiB of memory. Ratna was trained with public and synthetic speech, including machine-generated labels, and not on real call audio. NepTel 0.2 phone audio scores 67.32% WER, and exact phone numbers are not captured, so it should not be used for those tasks. This is a research release.

The full report is on the Ampixa blog.

Licence

If you use Ratna, or build on it, name it properly: Ratna (Ampixa Labs), in your model card or documentation. See LICENSE-MODEL.md.