Moonshine Streaming Medium

June 28, 2026 · View on GitHub

Useful Sensors' UsefulSensors/moonshine-streaming-medium ported to transcribe.cpp. A 245M-parameter encoder-decoder English ASR model designed for streaming use (ergodic encoder + sliding-window attention, 50 Hz time-domain frontend). Same family as the tiny and small variants; deepest of the three (14 / 14 layers) and widest hidden dims (encoder 768 / decoder 640).

What it's for

Offline English speech-to-text. The model takes a 16 kHz mono WAV and produces a transcript. It does not translate, has no multilingual capability, and does not emit timestamps.

See Useful Sensors' model card for training data, intended use, and upstream evaluation methodology.

Licensed MIT. Ported from upstream commit 57b8436, pinned 2026-05-06.

Download

QuantizationDownloadSizeWER (LibriSpeech test-clean)
F32moonshine-streaming-medium-F32.gguf1015 MB2.16%
F16moonshine-streaming-medium-F16.gguf509 MB2.16%
Q8_0moonshine-streaming-medium-Q8_0.gguf282 MB2.16%

WER is measured on the full LibriSpeech test-clean split (2620 utterances) with greedy decoding (num_beams=1, do_sample=False). F32 reference baseline: 2.16%. Quants are numerically indistinguishable from F32 on this manifest. Useful Sensors' self-reported number on this split is 2.08% from the Open ASR Leaderboard table; the +0.08pp residual matches the same scoring / text-normalization difference seen across the tiny and small variants (where the tiny cross-check against the HF Transformers reference on the same manifest landed within 0.01pp of our port), and is not a numerical drift in the port.

One utterance the model cannot end. A single LibriSpeech test-clean clip — 7176-92135-0020 (7.2 s; reference "DOUBLE NINE TWO THREE ELSINORE DOUBLE NINE YES HALLO IS THAT YOU HORATIO HAMLET SPEAKING") — drives the medium model into a hallucination loop: it collapses the repeated digits into an endless run of a single token and never emits end-of-stream. This is a property of the upstream weights, not the port — the HF Transformers reference does the identical thing (with max_new_tokens unbounded it emits an uninterrupted stream of 9s and never stops). The tiny and small variants terminate normally on this clip; only medium loops. Our decoder bounds the runaway with a duration-based generation budget (≈6.5 tokens per second of audio, plus a small floor — matching the model card's recommended max_new_tokens ≈ audio_seconds × 6.5), so the loop stops after a few dozen tokens instead of grinding to the decoder's position cap. The utterance is flagged via transcribe_was_truncated() and surfaced as an output truncated error (non-zero exit); its incomplete transcript is counted as a miss in the 2.16% WER above, so the baseline already includes this one pathological utterance.

Q6_K / Q5_K_M / Q4_K_M GGUFs are not currently shipped for this variant.

Quick Start

cmake -B build
cmake --build build

build/bin/transcribe-cli \
  -m models/moonshine-streaming-medium/moonshine-streaming-medium-Q8_0.gguf \
  samples/jfk.wav

If your audio is not already 16 kHz mono WAV, convert it first:

ffmpeg -i input.mp3 -ar 16000 -ac 1 output.wav

Performance

Cells are wall-clock latency (mean over 5 iterations after 2 warmups), with speedup over realtime in parentheses. Units: ms below 1 s, s above (2 decimal places).

Apple M4 Max

BackendSampleQ8_0
Metaljfk (11.0s)124 ms (89×)
Metaldots (35.3s)952 ms (37×)
CPUjfk (11.0s)281 ms (39×)
CPUdots (35.3s)1.11 s (32×)

macOS 26.4.1, transcribe.cpp 0d312ce.

AMD Ryzen 7 4750U Pro

BackendSampleQ8_0
Vulkanjfk (11.0s)570 ms (19×)
Vulkandots (35.3s)4.01 s (9×)
CPUjfk (11.0s)1.07 s (10×)
CPUdots (35.3s)6.50 s (5×)

Fedora 43, transcribe.cpp f243f34. Vulkan device: AMD Radeon Graphics (RADV RENOIR).

Benchmark reproduction:

uv run scripts/bench/run.py \
  --models moonshine-streaming-medium \
  --quants q8_0 \
  --samples jfk,dots \
  --backends metal,cpu,vulkan \
  --iters 5 --warmup 2 \
  --name moonshine-streaming-publication

Numerical Validation

transcribe.cpp is validated tensor-by-tensor against the HF Transformers reference (MoonshineStreamingForConditionalGeneration, fp32 inference, attn_implementation="eager") on samples/jfk.wav. All contract tensors fall within family tolerance, and the final transcript matches the reference. Last validated at commit 0d312ce.

FieldValue
ReferenceHF Transformers v5.7.0, UsefulSensors/moonshine-streaming-medium
Dump scriptscripts/dump_reference_moonshine_streaming_transformers.py
Manifesttests/golden/moonshine_streaming/moonshine-streaming-medium.manifest.json
Commanduv run scripts/validate.py all --family moonshine_streaming --variant moonshine-streaming-medium

Tolerances are recorded at family scope in tests/tolerances/moonshine_streaming.json. The dominant drift source is BLAS reduction-order differences between PyTorch's matmul kernels and ggml's mul_mat. Drift accumulates roughly linearly with depth across the 14-layer encoder, the adapter, and the 14-layer decoder; the family tolerances were widened from the tiny baseline using max(1.5 × observed, prior, 1e-6) to absorb the deeper-stack accumulation without inflating the budget where it isn't needed. The final logit budget remains well below 1e-3 absolute / 1e-4 mean.

Like the small variant, medium uses non-square encoder attention (encoder residual dim 768, attention dim 640). The C++ port carries both shapes through the encoder block.

Reproduction

Convert

uv run --project scripts/envs/moonshine_streaming \
  scripts/convert-moonshine_streaming.py UsefulSensors/moonshine-streaming-medium

Quantize

build/bin/transcribe-quantize \
  models/moonshine-streaming-medium/moonshine-streaming-medium-F32.gguf \
  models/moonshine-streaming-medium/moonshine-streaming-medium-F16.gguf \
  --quant F16

Validate

uv run scripts/validate.py all --family moonshine_streaming --variant moonshine-streaming-medium

WER sweep

uv run scripts/wer/run.py \
  --model models/moonshine-streaming-medium/moonshine-streaming-medium-F32.gguf \
  --manifest samples/wer/test-clean.manifest.jsonl \
  --out reports/wer/moonshine-streaming-medium-F32.test-clean.jsonl
uv run scripts/wer/score.py reports/wer/moonshine-streaming-medium-F32.test-clean.jsonl