moonshine-streaming-tiny-es-gguf-candidate

Moonshine Streaming Tiny (Spanish): candidate transcribe.cpp GGUF

Candidate Q8_0 GGUF conversion of moonshine-ai/moonshine-streaming-tiny-es for review by the transcribe.cpp project.

This is a review artifact, not a canonical transcribe.cpp release. Canonical models are published by maintainers under the handy-computer Hugging Face organization after numerical, WER, benchmark, and download-roundtrip validation.

Provenance

• Source repository: moonshine-ai/moonshine-streaming-tiny-es
• Source revision: 215dc49e20d2efcfc438cd788b02920639bd85a0
• Source format and dtype: model.safetensors, F32
• Architecture: moonshine_streaming encoder-decoder
• Language: Spanish (es)
• Parameters in the source checkpoint: approximately 27M
• Quantization: Q8_0 using transcribe-quantize

Candidate file

Compatibility adaptation

The Spanish checkpoint declares tie_word_embeddings=true and does not contain proj_out.weight. The Moonshine Streaming loader in transcribe.cpp 0.2.2 expects a separate dec.lm_head.weight and rejects tied-head metadata.

For compatibility, this candidate materializes the source model.decoder.embed_tokens.weight twice: once as dec.token_embd.weight and once as dec.lm_head.weight. The GGUF therefore truthfully describes its materialized layout as untied. Both tensors contain the same source values, so the resulting projection is mathematically equivalent to the tied source model. This costs approximately 4 MiB in the Q8_0 artifact but requires no runtime change.

Validation status

Completed:

• loads with transcribe-cpp 0.2.2;
• offline transcription smoke test;
• streaming feed/finalize smoke test using 500 ms chunks;
• local integration test with pi-transcribe.

Still required before canonical publication:

• tensor-level numerical validation against the pinned Transformers reference;
• Spanish WER evaluation on a representative acceptance manifest;
• canonical model-card and benchmark results;
• maintainer download-roundtrip validation.

Usage

The model expects 16 kHz mono PCM. With a transcribe.cpp checkout:

build/bin/transcribe-cli \
  -m moonshine-streaming-tiny-es-Q8_0.gguf \
  --stream-chunk-ms 500 \
  audio-16khz-mono.wav