Samuel Vitorino

[Samuel Vitorino @SamVitorino]

Released a new version of my side project: SoproTTS

A 135M parameter TTS model trained for ~$100 on 1 GPU, running ~20× real-time on a base MacBook M3 CPU.

v1.5 highlights (on CPU):

• 250 ms TTFA streaming latency
• 0.05 RTF (~20× real-time)
• Zero-shot voice cloning
• Smaller, faster, more stable

Still not perfect (OOD voices can be tricky, and there are still some artifacts), but a decent upgrade.

Open-source (Apache 2.0).

[Samuel Vitorino @SamVitorino]

@thsottiaux what is it with GPT-5.4/5.5 and “hand-waving”?

[Samuel Vitorino @SamVitorino]

One of the fastest voice cloning TTS models available, trained for just $100. Check it out:

[Samuel Vitorino @SamVitorino]

Released a new version of my side project: SoproTTS

A 135M parameter TTS model trained for ~$100 on 1 GPU, running ~20× real-time on a base MacBook M3 CPU.

v1.5 highlights (on CPU):

• 250 ms TTFA streaming latency
• 0.05 RTF (~20× real-time)
• Zero-shot voice cloning
• Smaller, faster, more stable

Still not perfect (OOD voices can be tricky, and there are still some artifacts), but a decent upgrade.

Open-source (Apache 2.0).

[Samuel Vitorino @SamVitorino]

Released a new version of my side project: SoproTTS

A 135M parameter TTS model trained for ~$100 on 1 GPU, running ~20× real-time on a base MacBook M3 CPU.

v1.5 highlights (on CPU):

• 250 ms TTFA streaming latency
• 0.05 RTF (~20× real-time)
• Zero-shot voice cloning
• Smaller, faster, more stable

Still not perfect (OOD voices can be tricky, and there are still some artifacts), but a decent upgrade.

Open-source (Apache 2.0).

[Samuel Vitorino @SamVitorino]

https://t.co/KT9moWMDBw

[Samuel Vitorino @SamVitorino]

Released a new version of my side project: SoproTTS

A 135M parameter TTS model trained for ~$100 on 1 GPU, running ~20× real-time on a base MacBook M3 CPU.

v1.5 highlights (on CPU):

• 250 ms TTFA streaming latency
• 0.05 RTF (~20× real-time)
• Zero-shot voice cloning
• Smaller, faster, more stable

Still not perfect (OOD voices can be tricky, and there are still some artifacts), but a decent upgrade.

Open-source (Apache 2.0).

[Samuel Vitorino @SamVitorino]

As a fun side project, I trained a small text-to-speech 169M parameter model that I call Sopro TTS.

169M parameters
Streaming support
Zero-shot voice cloning
0.25 RTF on CPU, meaning it generates 30 seconds of audio in 7.5 seconds
Requires 3-12 seconds of reference audio

[Samuel Vitorino @SamVitorino]

GitHub Repo
https://t.co/KT9moWMDBw

[Samuel Vitorino @SamVitorino]

As a fun side project, I trained a small text-to-speech 169M parameter model that I call Sopro TTS.

169M parameters
Streaming support
Zero-shot voice cloning
0.25 RTF on CPU, meaning it generates 30 seconds of audio in 7.5 seconds
Requires 3-12 seconds of reference audio

[Samuel Vitorino @SamVitorino]

Inspired by @karpathy's llama2.c and as a fun learning weekend project, I decided to write the most simple Rust code to run inference on the new Gemma 2b models. One moment you are writing some matmuls, the other you are talking to your computer. https://t.co/iwHZMsr5BY