v0.8.0: Multi-model TTS with Kokoro, Bark, Parler, XTTS

What's New

Narro now supports 6 TTS models behind one API. Pull, serve, switch between them with a single model field.

New Models

Model	Size	Speed	Key Feature
Kokoro 82M	200 MB	~5-7x real-time	54 voices, 6 languages, excellent quality
Parler-TTS	600 MB	Medium	Describe voice in natural language
XTTS v2	1.8 GB	Slow	Zero-shot voice cloning from 6s audio
Bark	1.5 GB	Slow	130 presets, 13 languages
Bark Small	600 MB	Slow	Lighter Bark
Soprano 80M	374 MB	Real-time	Streaming, CPU-friendly

Auto-install dependencies on pull

narro models pull kokoro-82m    # auto: pip install kokoro soundfile misaki[en]
                                # warns: sudo apt install espeak-ng

Each model declares its pip_extras and system_packages. narro models pull installs pip packages automatically and warns about missing system packages.

Other improvements

• Instant CLI startup: deferred heavy imports (torch/transformers), narro --help in <100ms
• bark-small gets distinct model_id: no longer collides with bark
• Voice list display: shows count instead of dumping 130 names
• Optional dependency groups in pyproject.toml: [kokoro], [parler], [xtts], [all]

Verified Kokoro performance

af_heart:   5.6s audio in 2.2s (2.5x real-time)
am_adam:    ~2s audio in 0.5s  (~5x real-time)
bm_george:  ~2s audio in 0.4s  (~7x real-time)