SteerTTS edge
The edge version of SteerTTS, with a size of only 11.84MB, aimed at mobile and embedded devices and supporting controllable speech synthesis.
Timeline (last 90 days)
WhissleAI released the edge version of SteerTTS (SteerTTS edge), a steerable English speech synthesis model whose entire distribution package is only 11.84 MB.
SteerTTS edge's 11.84 MB includes model weights, voices, text frontend, and the Python runtime, which is everything the application needs to distribute.
SteerTTS edge's acoustic model has 9.25M parameters, uses an attention + bidirectional Mamba-3 architecture, and is compressed to int4 through quantization-aware training (QAT).
SteerTTS edge's vocoder is a 6.2M-parameter Vocos-style model distilled from charactr/vocos-mel-24khz and subjected to int4 QAT.
SteerTTS edge's text frontend is instead implemented purely in numpy, and the runtime no longer depends on NLTK, g2p_en, torchaudio, or num2words.
The authors state that SteerTTS edge's roughly 0.1 UTMOS loss compared with v1 comes almost entirely from the small int4 vocoder.
After quantization, SteerTTS edge retains 12 control dimensions (speaking rate, loudness, pitch, pitch range, sentence-end rise, pause, tilt, breathiness, etc.), each retaining at least 95% of the fp32 version's control effect.
SteerTTS edge supports 50 style tags and inline markup (such as [pause 400ms], <slow>, <whisper>), allowing pauses, speaking rate, and whisper style to be controlled directly in the text.
In a pure CPU environment (2018 Threadripper, under load), SteerTTS edge achieves an RTF of 0.053 with 4 threads (about 19x real-time), with first-packet audio latency of about 120 ms thanks to sentence-level streaming.
SteerTTS edge supports only English and includes 200 built-in voices (compared with 2,825 in the full version).
3 more timeline events
All Facts (13)
Most of SteerTTS edge's emotion tags are just style switches, and the fast and anger categories are relatively weak.
50%UnverifiedSteerTTS edge's 11.84 MB includes model weights, voices, text frontend, and the Python runtime, which is everything the application needs to distribute.
50%UnverifiedSteerTTS edge's acoustic model has 9.25M parameters, uses an attention + bidirectional Mamba-3 architecture, and is compressed to int4 through quantization-aware training (QAT).
50%UnverifiedSteerTTS edge's vocoder is a 6.2M-parameter Vocos-style model distilled from charactr/vocos-mel-24khz and subjected to int4 QAT.
50%UnverifiedSteerTTS edge's text frontend is instead implemented purely in numpy, and the runtime no longer depends on NLTK, g2p_en, torchaudio, or num2words.
50%UnverifiedThe authors state that SteerTTS edge's roughly 0.1 UTMOS loss compared with v1 comes almost entirely from the small int4 vocoder.
50%UnverifiedAfter quantization, SteerTTS edge retains 12 control dimensions (speaking rate, loudness, pitch, pitch range, sentence-end rise, pause, tilt, breathiness, etc.), each retaining at least 95% of the fp32 version's control effect.
50%UnverifiedSteerTTS edge supports 50 style tags and inline markup (such as [pause 400ms], <slow>, <whisper>), allowing pauses, speaking rate, and whisper style to be controlled directly in the text.
50%UnverifiedIn a pure CPU environment (2018 Threadripper, under load), SteerTTS edge achieves an RTF of 0.053 with 4 threads (about 19x real-time), with first-packet audio latency of about 120 ms thanks to sentence-level streaming.
50%UnverifiedSteerTTS edge supports only English and includes 200 built-in voices (compared with 2,825 in the full version).
50%UnverifiedSteerTTS edge does not support voice cloning, free-text prompting, or the flow refiner on-device; these capabilities are retained in the full model.
50%UnverifiedSteerTTS edge's weights are dequantized to fp32 on load, so this version saves size rather than inference speed, unless operator kernels are rewritten for edge devices.
50%UnverifiedWhissleAI released the edge version of SteerTTS (SteerTTS edge), a steerable English speech synthesis model whose entire distribution package is only 11.84 MB.
50%