Arabic streaming speech recognition, 27.0M parameters. Same
architecture as
moonshine-ai/moonshine-streaming-tiny,
trained for Arabic with a 12,288-entry Arabic tokenizer.
Moonshine Streaming pairs a 50 Hz time-domain audio frontend with a
sliding-window Transformer encoder, so it transcribes incrementally rather than
waiting for an utterance to finish. It is intended for on-device use on
edge-class hardware.
Checkpoint identity
This repository is a conversion of one specific training checkpoint, recorded
here because the weights behind a language move as later stages win:
Checkpoint
ar12k_tiny_stageC_best.safetensors
Stage
C (read-speech mix)
Architecture
slinkier_prime_adapted
Tokenizer
tokenizer_ar12k.json, vocab 12,288
Snapshot taken
2026-08-24
Parameters
27.0M
If you need reproducibility, pin the revision of this repository rather than
tracking main.
1from transformers import MoonshineStreamingForConditionalGeneration, AutoProcessor
2import torch
34model = MoonshineStreamingForConditionalGeneration.from_pretrained(5"moonshine-ai/moonshine-streaming-tiny-ar"6).eval()7processor = AutoProcessor.from_pretrained("moonshine-ai/moonshine-streaming-tiny-ar")89inputs = processor(audio, return_tensors="pt", sampling_rate=16000)1011# Cap the output length. Like other seq2seq ASR models this one can fall into a12# repetition loop, and short or noisy clips are where it happens.13seq_lens = inputs.attention_mask.sum(dim=-1)14max_new_tokens =int((seq_lens *6.5/16000).max().item())+21516generated = model.generate(**inputs, max_new_tokens=max_new_tokens)17print(processor.batch_decode(generated, skip_special_tokens=True)[0])
Pass the attention_mask. The encoder applies its per-layer sliding windows
only when it is given one; called without a mask it attends over the whole
utterance instead, which is a different model from the one that was trained. The
processor returns the mask, so the snippet above is the safe form. The processor
also pads audio to a whole number of 80-sample frames, which the frontend
requires.
Architecture
Encoder
6 layers, width 320, 8 heads, sliding windows (16, 4) on the first two and last two layers and (16, 0) between
Decoder
6 layers, width 320, 8 heads, RoPE over 32 of each head's 40 dimensions
Frontend
50 Hz features, CMVN, asinh compression, two causal stride-2 convolutions
Adapter
learned absolute positional embeddings before the decoder
The lookahead layers give roughly 80 ms of lookahead; the intermediate layers
have none.
Training data
Trained on a large-scale automatically labeled Arabic corpus:
Crawled corpus, roughly 10,000 hours, pseudo-labeled and unaudited.
The crawled transcripts are pseudo-labels: they were produced by running a
Whisper-family teacher model over crawled audio, not by human transcription. The
model therefore inherits the teacher's error modes, including its handling of
proper nouns, numerals and code-switching. No human-verified transcript was used
for the bulk of training.
Evaluation
Arabic is scored on word error rate (WER), after the usual case and
punctuation normalization. Mandarin and Japanese in this model family are
instead scored on no-space CER, because they are written without spaces; every
other language, this one included, uses WER.
suite_ar is Common Voice Arabic and FLEURS Arabic. The FLEURS panel is
Egyptian Arabic; Common Voice is broader but dominated by very short clips.
Modern Standard Arabic and the regional dialects are not measured separately,
and no dialect other than Egyptian is represented in the read-speech panel.
Do not quote a full-panel Arabic number from this card. The figures here are
a seeded 400-clip sample. A wider Arabic measurement in our own notes reads
15.205, but it was taken on a 2,500-row draw of the 10,480-row Common Voice
panel, so it is not a full-coverage result either. The numbers below are sound
as relative measurements between these three builds, which is what they are for.
Seeded 400-utterance sample, batch 1
Batch 1 is the honest number for deployment. Batched evaluation zero-pads short
clips up to the longest in the batch, and that trailing silence flatters the
model.
Panel
WER
cv_ar
17.91
fleurs_ar
12.56
macro
15.231
This repository against the training checkpoint
These weights were converted from the neo training checkpoint, and the
conversion was checked by measurement rather than inspection: same seeded
sample, same batch size, same normalizer. A conversion that loads and emits
plausible text can still have a permuted weight mapping, which only a score
catches.
cv_ar
fleurs_ar
macro
Training checkpoint
17.91
12.56
15.231
This repository
17.86
12.56
15.207
399/400 and 400/400 transcripts are byte-identical.
The quantized build we ship
The .ort package served to the Moonshine deployment library is quantized to
int8 from these same weights, and scores 15.533 against 15.231 for the float
checkpoint on the same sample under the same stopping rule -- a cost of +0.302
WER. That build is a different artifact from this repository, which is float32.
Limitations
Machine-labeled training data. See above; the model reproduces its
teacher's mistakes as well as its strengths.
Repetition loops on short clips. Like other seq2seq ASR models this one
can fall into a repetition loop, and short or noisy clips are where it
happens. Cap the output length, as the usage snippet does.
Evaluated on 2 panels only. No evaluation of telephony, children's
speech, heavy dialect, or noisy far-field conditions.
Short-clip sensitivity. The Common Voice panel is 93% short clips, and
short clips are where this architecture's stopping decision is weakest; int8
quantization costs four times as much on that panel as on FLEURS.
Out-of-scope use
Not intended for non-consensual surveillance, speaker identification, or
high-stakes decisions.
