Add Moonshine ASR backend (CPU-friendly, with Arabic-tuned default)

[Ahmed-Ezzat20]

Summary

Adds a sixth transcription backend, Moonshine (Useful Sensors), via the
Hugging Face Transformers integration. Moonshine is a small (27–61 M params),
CPU-friendly ASR model and the Arabic-specialised checkpoint
moonshine-tiny-ar
matches whisper-medium quality on Arabic Fleurs/CommonVoice at 28× fewer
parameters — a great fit for users who want a free, local, Arabic-strong
alternative to Deepgram or full Whisper.

pip install \"mazinger[transcribe-moonshine]\"
mazinger transcribe audio.mp3 --method moonshine --language ar -o subs.srt

The default model is picked from --language: moonshine-tiny-ar for Arabic
and moonshine-base otherwise. Users can override with --model.

Implementation notes

• _transcribe_moonshine() in mazinger/transcribe.py: loads
  MoonshineForConditionalGeneration + AutoProcessor, runs Silero VAD
  ourselves (Moonshine has no native long-audio chunker and was trained on
  ≤30 s clips), then transcribes each VAD chunk and stamps it with the chunk's
  start/end times.
• Hallucination guard: applies the model card's 13 tokens/sec cap as
  max_length per chunk to suppress runaway outputs on near-silent regions.
• Model caching: reuses the existing _whisper_cache, so
  transcribe.clear_cache() correctly frees both Moonshine and Whisper models
  between runs.
• Audio preprocessing: reuses _preprocess_audio() so a project that
  switches between faster-whisper and moonshine doesn't re-encode the audio.

Cross-cutting fix triggered by Moonshine's hallucination behaviour

While testing on a real Arabic clip with a music outro, Moonshine's third VAD
chunk got stuck in a \"كما قلت، كما قلت، كما قلت...\" loop. The existing
_REPEATED_WORD_RE only collapses whitespace-separated repeats, so
punctuation-separated stuck-token loops slipped through.

Added _REPEATED_PHRASE_RE that collapses 3+ phrase repeats joined by Latin
or Arabic punctuation. This benefits every backend that can produce
stuck-token loops on near-silent / OOD audio — Whisper, faster-whisper, and
WhisperX all do this occasionally too. Verified with 7 positive cases (Arabic
multi-word, Latin multi-word, single-word + punct, mixed real text + loop)
and a negative case ensuring normal sentences with shared words like "and"
are not over-collapsed.

Tests

tests/test_moonshine.py — 9 unit tests covering default-model selection,
the new phrase-cleanup regex (positive + negative), method-literal
membership, and the dispatch error message. They run without downloading
the model so they're cheap to add to CI. The full test suite (14 tests
including the existing MLX ones) passes.

End-to-end test result

Tested on the same 58-second Arabic clip used to validate Deepgram earlier:

Run	Wall-clock	Real-time factor	Notes
Cold (downloads model)	122 s	~0.5×	Surfaced the punctuation-loop bug above
Warm (cached model)	24 s	~2.4×	Loop collapsed correctly, two real-content chunks transcribe cleanly

Quality is comparable to Deepgram on the speech portions; the local zero-cost
zero-network properties are the main win.

Files

File	Change
mazinger/transcribe.py	New _transcribe_moonshine(), default-model picker, dispatch case, docstring updates, new _REPEATED_PHRASE_RE
mazinger/cli/_transcribe.py	--method moonshine choice + help text
mazinger/cli/_groups.py	--transcribe-method moonshine choice + help text
pyproject.toml	New transcribe-moonshine extra (transformers>=4.49, torch>=2.4, silero-vad>=5.0)
tests/test_moonshine.py	9 new unit tests
README.md	Feature list, install extras, dedicated Quick Start section
docs/installation.md	Local transcription extras + task matrix row
docs/cli-reference.md	Choice lists + transcribe example
docs/quick-start.md	Moonshine usage example