moonshine-ai/moonshine-streaming-tiny-tl

Moonshine Streaming Tiny — Tagalog

Tagalog streaming speech recognition, 27.0M parameters. Same architecture as moonshine-ai/moonshine-streaming-tiny, trained for Tagalog with a 12,288-entry Tagalog 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	tl12k_tiny_stageA_best.safetensors
Stage	A (alignment) -- still training, see below
Architecture	slinkier_prime_adapted
Tokenizer	tokenizer_tl12k.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.

This is a snapshot of a run that had not finished. The Tagalog Stage A run was still training when these weights were taken: its best score was 16.600 and it had been flat around 17.16 for 31 consecutive readings near the end of its epochs, so this checkpoint is the best available rather than a converged one. It was published deliberately, and a later refresh is expected.

Usage

pip install --upgrade transformers datasets[audio]

from transformers import MoonshineStreamingForConditionalGeneration, AutoProcessor
import torch

model = MoonshineStreamingForConditionalGeneration.from_pretrained(
    "moonshine-ai/moonshine-streaming-tiny-tl"
).eval()
processor = AutoProcessor.from_pretrained("moonshine-ai/moonshine-streaming-tiny-tl")

inputs = processor(audio, return_tensors="pt", sampling_rate=16000)

# Cap the output length. Like other seq2seq ASR models this one can fall into a
# repetition loop, and short or noisy clips are where it happens.
seq_lens = inputs.attention_mask.sum(dim=-1)
max_new_tokens = int((seq_lens * 6.5 / 16000).max().item()) + 2

generated = model.generate(**inputs, max_new_tokens=max_new_tokens)
print(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 Tagalog corpus:

• YouTube crawl, roughly 163,300 hours, pseudo-labeled.
• Podcast crawl, roughly 3,300 hours, pseudo-labeled.

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

Tagalog 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_tl is FLEURS Tagalog, and it is the only Tagalog panel we hold. The headline number is therefore one read-speech set rather than a macro over two, and it does not measure spontaneous speech or Taglish code-switching, both of which are heavily represented in the training data.

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
fleurs_tl	15.05

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.

	fleurs_tl	macro
Training checkpoint	15.05	15.046
This repository	14.91	14.905

399/400 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 14.867 against 15.046 for the float checkpoint on the same sample under the same stopping rule -- a difference of -0.179, which is inside the noise of a 400-clip sample and should not be read as the quantized build being better or worse. 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 1 panel only. No evaluation of telephony, children's speech, heavy dialect, or noisy far-field conditions.
• Snapshot of an unfinished run, scored on one panel. Both of these make the headline number less trustworthy than the same number would be for the other languages here.

Out-of-scope use

Not intended for non-consensual surveillance, speaker identification, or high-stakes decisions.

License

MIT.