Skip to content

Add Sopro V2 Turbo support via a CPU speech backend #36

Open

Description

@pirumpi

Summary

Research date: 2026-09-06. Add a scoped, initially experimental Sopro V2 Turbo CPU speech backend and qualify it before public catalog activation. This is an integration candidate, not current Tapioca support.

Official sources, release and access

Weights and runtime are Apache-2.0; the checkpoint is public and ungated. Preserve attribution/license files. Reference recordings need appropriate rights and consent; no built-in watermark is provided. The product must not imply that downloadable weights authorize impersonating another person.

Architecture and hardware guidance

120M-class TTS combining an autoregressive semantic-token transformer, flow-matching acoustic head, semantic/speaker encoders and causal vocoder. Languages: English, European Portuguese, French, German. Output: 24 kHz.

Upstream publishes single-stream M3 CPU figures of about 300 ms first audio and 0.21 streaming / 0.24 offline real-time factor. Those are upstream measurements, not Tapioca guarantees. This gives a documented consumer-local CPU path without requiring a GPU. Minimum RAM and Windows/Linux latency remain qualification outputs and must be measured before publishing catalog guidance.

Exact downloadable bundle

Repository samuel-vitorino/sopro-v2-turbo, revision f747f9edfb7b0233a3b7105af3a75603a7213d26. All six exact revision URLs responded HTTP 200 today; no full-weight download or inference test was performed in this audit. Total: 634,337,563 bytes (~605 MiB).

Artifact Bytes SHA-256
config.json 4,482 b8593ff96b2976215ff64e58e784f4d667f4f0f205ae8686b57e78c10f42d6d7
model.safetensors 486,332,212 5ddc2e905ce68e015b3760f9437969f6fd7ef0af601b8e457fc18344bb908ffa
semantic_encoder.safetensors 82,497,708 d1d1bccdfffdcd6d01d802a3e5dbefa0674c660f5022d87ec14e3271596c5c5f
speaker_encoder.safetensors 10,784,100 67422a45d801542ee0aa55ce824f570ceb1c6b2283b4e74978c1b4673b49acd7
tokenizer.model 370,821 2d76e7a4e8dbd0a4d2137c13200ac9710ad3fd93fad20a61f71102ab7bae754d
vocoder.safetensors 54,348,240 07e5561491cce41f7f90cfdb94b2ff263ff5742c3d89339db99b17ad82cc3f44

Large-file hashes are Hub LFS metadata; the config digest was computed from its downloaded bytes. Verify every hash during the actual integration download.

Runtime and platform scope

Pin sopro==2.2.0 or the above exact official commit: current artifacts require at least 2.2.0, and older package/checkpoint combinations are incompatible. Resolve exact Python/PyTorch/torchaudio and supporting dependencies in an isolated runtime rather than changing existing speech environments.

The official artifact loader accepts a local directory containing the six files, allowing a fully offline-after-download implementation. The inference API provides synthesis and streaming.

Initial target: a dedicated speech-sopro PyTorch CPU adapter. Qualify macOS arm64, Windows x64 and Linux x64 separately. Upstream CUDA/optional MPS support is outside the initial advertised scope. No llama.cpp, MLX, Diffusers, ONNX Runtime/DirectML or Vulkan compatibility is claimed. Upstream browser ONNX is a separate runtime/project, not reuse of Tapioca's ONNX image adapter.

Implementation and acceptance tests

  1. Add isolated backend dispatch, a pinned dependency environment and exact artifact manifest; keep the candidate out of the public picker until qualification passes.
  2. Use the local-directory CPU API with a consented 5–20-second reference; keep preprocessing local and propagate explicit voice consent.
  3. Reuse atomic incremental PCM16 WAV output patterns; implement cancellation cleanup and validate language/seed options.
  • Clean download verifies every artifact; missing/corrupt files return clear errors.
  • Offline-after-cache generation performs no hidden component download.
  • All four languages produce complete intelligible short and long speech.
  • Valid, silent, malformed, short and long references have tested handling; missing consent blocks custom cloning.
  • Chunk ordering, nonempty audio, final 24 kHz WAV integrity, cancellation and atomic completion pass.
  • Test numbers/symbols, normalization and mixed-language limitations.
  • Measure load time, peak RSS, first-audio latency and real-time factor on each target OS; publish honest RAM guidance.
  • Streaming and offline are evaluated independently, not required to be bit-identical.
  • Do not advertise live desktop playback unless that feature is separately implemented/tested.
  • Existing speech/catalog/adapter and desktop tests stay green.

Risks and blockers

No unresolved upstream inference blocker was identified, but Tapioca integration and platform/model-inference qualification have not run. Voice-reference quality, normalization and mixed-language pronunciation can vary. Package/checkpoint drift is a concrete compatibility risk; use the current exact pair. CPU performance claims must be reproduced, and model download size is not a RAM estimate. This is not a swap-in artifact for Audio8, Pocket, Qwen or Chatterbox.

Reddit field evidence — non-authoritative

Author reports about reference-distortion corrections reinforce pinning current artifacts. These posts are discovery/field signals, not independent proof of Tapioca performance.

Duplicate check

Searched all open and closed issues for Sopro, model name, CPU speech/streaming and likely title, and inspected #30/#31 plus their September 5 comments. They track different models and blockers; no Sopro issue exists. No routine updates to those issues or historical #29 are needed.

Activity

  1. pirumpi commented on Sep 18, 2026

    @pirumpi
    ContributorAuthor

    Experimental implementation is in draft PR #44: six SHA-256-pinned artifacts, isolated sopro==2.2.0 CPU adapter, local offline streaming-to-WAV, explicit reference consent, and opt-in Mac hardware test. An end-to-end Tapioca pull/tts smoke passed on Apple Silicon with a synthetic reference (24 kHz WAV, 1.13 s first audio after environment setup). I am keeping this issue open until multilingual quality, reference handling, cancellation, memory, and Windows/Linux CPU qualification pass; the candidate is not in the public catalog.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions