kala-tts 0.1.0

CPU-native Nepali TTS — real_nepali G2P + VITS ONNX inference

kala-tts

CPU-native Nepali text-to-speech. The first open-source Nepali TTS built without eSpeak.

pip install kala-tts

import kala_tts

kala_tts.synthesize_to_file("नमस्कार, कसरी हुनुहुन्छ?", "out.wav")

What makes it different

eSpeak-ng maps Nepali affricates to alveolar labels (ts, tsh) that do not match how Kathmandu speakers actually produce च and छ. kala-tts uses the real_nepali G2P — built from Khatiwada (2009) and a 48 000-entry curated lexicon — which outputs palatal ch/chh as the acoustic target.

Feature	eSpeak ne	kala-tts
च / छ	ts / tsh (alveolar)	ch / chh (palatal)
Gemination	often lost	explicit ː tokens
Schwa deletion	heuristic	rule-based, audited
Latin code-switch	undefined	letter-by-letter + lexicon
Lexicon	none	48 000 entries

API

import kala_tts

# Returns WAV bytes (16-bit PCM mono, 22050 Hz)
wav: bytes = kala_tts.synthesize("नेपाल सुन्दर देश हो।", speaker="kala")

# Write directly to file
kala_tts.synthesize_to_file("नमस्कार", "output.wav", speaker="barsha")

# Speed control
wav = kala_tts.synthesize("राम्रो दिन!", speed=0.85)  # slower

# List speakers
print(kala_tts.list_speakers())
# ('kala', 'barsha', 'slr143_F', 'slr43_0546', 'slr43_2099')

CLI

kala-tts "नमस्कार, कसरी हुनुहुन्छ?" --speaker kala -o out.wav
kala-tts --list-speakers
echo "नेपाल सुन्दर देश हो।" | kala-tts -o out.wav

Speakers

Name	Data	Notes
kala	human studio	recommended
barsha	human recording	most data
slr143_F	OpenSLR-143	neutral prosody
slr43_0546	OpenSLR-43	read speech
slr43_2099	OpenSLR-43	read speech

Model

The first call auto-downloads the ONNX model (~60 MB) from ampixa/real-nepali-v0.2-kala via huggingface_hub and caches it locally. No internet needed after first run. No GPU required.

• Architecture: VITS (ONNX FP32)
• Sample rate: 22050 Hz
• CPU RTF: ~0.020 (50× real-time on a laptop)

License

MIT (code) — CC-BY-SA 4.0 (model weights and training data). See LICENSES.md for full attribution.