shama-tts v0.1

shama-tts is a small, fast text-to-speech model from SuperHQ Labs. You can put expression tags such as [laugh] inline in the text. The initial release supports English.

• Parameters: 14.8M (acoustic model)
• Architecture: Matcha-TTS (flow matching) with the Vocos vocoder
• Output: 24 kHz mono
• Languages: English
• Voices: 1 (heart); slots for 16 are reserved for later versions
• Speed: about 7× faster than real time on an Apple M1 CPU (4 threads, 10 steps)
• License: Apache-2.0

Try it: playground

Expression tags

Put a tag where the sound should happen:

Wait, you actually ate the whole cake? [laugh] That's honestly impressive.

Tag	Sound
[laugh]	a short laugh
[sigh]	a sigh
[gasp]	a gasp

Usage

hf download superhq-ai/shama-tts --local-dir shama-tts
pip install ./shama-tts

import soundfile as sf
from shama_tts import ShamaTTS

tts = ShamaTTS.from_pretrained("shama-tts")          # or "superhq-ai/shama-tts"
wav = tts("Oh no. [sigh] I left the keys on the kitchen table again.",
          speed=1.0, temperature=0.667, steps=10, seed=0)
sf.write("out.wav", wav, tts.sample_rate)

• Long text: it is split into sentences automatically.
• temperature: lower values give a steadier delivery, higher values a more varied one.
• steps: the number of flow-matching steps. Use 4 for speed and 10–24 for quality.
• In the browser: onnx/ holds an ONNX export (encoder, one flow-matching step, Vocos) that the playground runs with ONNX Runtime Web.

Training data

v0.1 was trained on about 14 hours of synthetic English speech. The speech was made with permissively licensed open-source TTS models (Apache-2.0 and MIT) reading public-domain text.

Evaluation

Metric	Value
Word error rate on 100 held-out sentences (Whisper-small)	2.6%

Limitations

• The initial release has one voice and supports English.
• Only the three tags above are supported; unknown tags raise an error.
• Questions often end with falling intonation, like statements. This is most noticeable when there's no question word ("You're coming tonight?").
• Text is spoken one sentence at a time. Sentences over about 220 characters are split at a comma or space, which can add an odd pause.
• Rare words and names follow the misaki G2P and may be mispronounced.

Responsible use

Generated audio is not watermarked. Don't use this model to impersonate real people, to make deceptive content, or to get around consent.

License

• Weights and inference code: Apache-2.0.
• Vendored Matcha-TTS modules: shama_tts/matcha/ contains inference modules from Matcha-TTS, MIT-licensed (© 2023 Shivam Mehta). See shama_tts/matcha/LICENSE.
• Vocoder: Vocos (charactr/vocos-mel-24khz) is MIT-licensed.