Post
69
LinnetTTS: a 41 M-parameter English text-to-speech model with voice prompts that streams from the first 80 ms and runs in real time on one phone CPU core. Code Apache-2.0; weights under the Linnet Weights License (free to use in your own products, commercial included).
Same machine, same three texts, each model through its own Python package. Speed only, quality not measured:
- Linnet 41 M: RTF 10.6 % on one core (3.4 % on four), first audio 87 ms, 187 MB RSS, voice prompts, streams every 80 ms
- pocket-tts 100 M: 23.8 % on two cores, 968 MB, voice prompts, streams every 80 ms
- Kokoro-82M: 51.7 %, 1.2 GB, whole utterance per call, 8 languages
- Piper medium 16 M: 5.8 %, 503 MB, no voice prompts, per-sentence chunks
On phones: iPhone 13 Pro Max 5.5 ms per 80 ms step, iPhone 11 7 ms, Snapdragon 662 14 ms on four cores.
š§ Hear the same sentence from all four, side by side: the "Compare" tab of the demo Space VoxRT/linnet-tts (type your own text for Linnet there too)
š¦ Weights and model card: VoxRT/linnet-en-35m
š¦ Code, Rust runtime, bindings for Linux C, Python (pip install linnet-tts), Android and iOS, and the benchmark script: https://github.com/VoxRT/LinnetTTS
š Method: https://github.com/VoxRT/LinnetTTS/blob/main/docs/benchmarks.md
š Write-up with the trade-offs and the samples: https://huggingface.co/blog/VoxRT/small-cpu-text-to-speech-models-compared-piper-kok
š Collection of small CPU TTS models and their demos: VoxRT/on-device-tts-small-models-that-run-on-a-cpu-6aec390f2f84a4407d227cf4
English only for now; no text normalisation, so write numbers as words. If it is useful to you, a like on the model or the Space helps other people find it. Questions welcome.
Same machine, same three texts, each model through its own Python package. Speed only, quality not measured:
- Linnet 41 M: RTF 10.6 % on one core (3.4 % on four), first audio 87 ms, 187 MB RSS, voice prompts, streams every 80 ms
- pocket-tts 100 M: 23.8 % on two cores, 968 MB, voice prompts, streams every 80 ms
- Kokoro-82M: 51.7 %, 1.2 GB, whole utterance per call, 8 languages
- Piper medium 16 M: 5.8 %, 503 MB, no voice prompts, per-sentence chunks
On phones: iPhone 13 Pro Max 5.5 ms per 80 ms step, iPhone 11 7 ms, Snapdragon 662 14 ms on four cores.
š§ Hear the same sentence from all four, side by side: the "Compare" tab of the demo Space VoxRT/linnet-tts (type your own text for Linnet there too)
š¦ Weights and model card: VoxRT/linnet-en-35m
š¦ Code, Rust runtime, bindings for Linux C, Python (pip install linnet-tts), Android and iOS, and the benchmark script: https://github.com/VoxRT/LinnetTTS
š Method: https://github.com/VoxRT/LinnetTTS/blob/main/docs/benchmarks.md
š Write-up with the trade-offs and the samples: https://huggingface.co/blog/VoxRT/small-cpu-text-to-speech-models-compared-piper-kok
š Collection of small CPU TTS models and their demos: VoxRT/on-device-tts-small-models-that-run-on-a-cpu-6aec390f2f84a4407d227cf4
English only for now; no text normalisation, so write numbers as words. If it is useful to you, a like on the model or the Space helps other people find it. Questions welcome.