SteerTTS-edge
An on-device text-to-speech (TTS) model of approximately 12MB that supports 200 voices and 50 styles. Through quantization-aware training it compresses the acoustic model and vocoder to int4 precision, and its text front end is implemented purely in numpy.
Timeline (last 90 days)
SteerTTS-edge has a total size of 11.84MB, including model weights, voice data, text frontend, and the Python runtime.
SteerTTS-edge's acoustic model is a 9.25-million-parameter attention + bidirectional Mamba-3 architecture, compressed to int4 precision through quantization-aware training (QAT).
SteerTTS-edge's vocoder is a 6.2-million-parameter Vocos-style model distilled from charactr/vocos-mel-24khz, also using int4 QAT.
SteerTTS-edge's text frontend is implemented purely in numpy, requiring no NLTK, g2p_en, torchaudio, or num2words at runtime.
Under local evaluation following the Open TTS Leaderboard protocol (not an official submission), SteerTTS-edge's v1 fp32 version achieves a WER of 0.96% and a UTMOS of 4.20 on Seed-TTS.
After quantization, SteerTTS-edge retains at least 95% of the fp32 version's control effect across 12 numerical controls (speaking rate, loudness, pitch, pitch range, sentence-final rise, pause, tilt, breathiness, etc.).
SteerTTS-edge offers 50 style tags and inline markup syntax, supporting effects such as pauses, speaking rate, and whispers embedded directly in the text stream.
On a 2018 Threadripper CPU (under load) with 4 threads, SteerTTS-edge achieves an RTF of 0.053, about 19x real-time speed, with first-packet audio of about 120ms, supporting sentence-level streaming output.
SteerTTS-edge's weights are dequantized back to fp32 on load, so the benefit of int4 quantization is size reduction rather than inference speedup.
SteerTTS-edge supports only English and offers 200 voices (a subset of the full model's 2,825 voices), with no voice cloning, free-text prompting, or streaming refiner.
1 more timeline events
All Facts (11)
SteerTTS-edge has a total size of 11.84MB, including model weights, voice data, text frontend, and the Python runtime.
50%UnverifiedSteerTTS-edge's acoustic model is a 9.25-million-parameter attention + bidirectional Mamba-3 architecture, compressed to int4 precision through quantization-aware training (QAT).
50%UnverifiedSteerTTS-edge's vocoder is a 6.2-million-parameter Vocos-style model distilled from charactr/vocos-mel-24khz, also using int4 QAT.
50%UnverifiedSteerTTS-edge's text frontend is implemented purely in numpy, requiring no NLTK, g2p_en, torchaudio, or num2words at runtime.
50%UnverifiedUnder local evaluation following the Open TTS Leaderboard protocol (not an official submission), SteerTTS-edge's v1 fp32 version achieves a WER of 0.96% and a UTMOS of 4.20 on Seed-TTS.
50%UnverifiedAfter quantization, SteerTTS-edge retains at least 95% of the fp32 version's control effect across 12 numerical controls (speaking rate, loudness, pitch, pitch range, sentence-final rise, pause, tilt, breathiness, etc.).
50%UnverifiedSteerTTS-edge offers 50 style tags and inline markup syntax, supporting effects such as pauses, speaking rate, and whispers embedded directly in the text stream.
50%UnverifiedOn a 2018 Threadripper CPU (under load) with 4 threads, SteerTTS-edge achieves an RTF of 0.053, about 19x real-time speed, with first-packet audio of about 120ms, supporting sentence-level streaming output.
50%UnverifiedSteerTTS-edge's weights are dequantized back to fp32 on load, so the benefit of int4 quantization is size reduction rather than inference speedup.
50%UnverifiedSteerTTS-edge supports only English and offers 200 voices (a subset of the full model's 2,825 voices), with no voice cloning, free-text prompting, or streaming refiner.
50%UnverifiedMost of SteerTTS-edge's emotion tags are just style switches, and the fast and anger categories are relatively weak.
50%