The truly free, open-source, local-first, cross-platform CapCut replacement. Scriptable through a JSON-RPC API & MCP.
Rust ffmpegfree-video-editormodel-context-protocolnon-linear-editoroffline-firstopen-source-video-editorself-hostedvideo-editingvideo-editing-automationvideo-editing-softwarevideo-editing-toolsvideo-editorvoice-cloningyoutube-automationyoutube-shorts
[package]
name = "concat-speech"
version.workspace = true
edition.workspace = true
rust-version.workspace = true
license.workspace = true
description = "Speech in and out: on-device transcription with whisper.cpp, and text to speech with Kokoro."
[dependencies]
concat-host.workspace = true
concat-media.workspace = true
serde = { version = "1", features = ["derive"] }
serde_json.workspace = true
# Model downloads. Blocking on purpose: downloads run on their own thread,
# and an async HTTP stack would be a second runtime's worth of dependencies
# for one GET.
ureq = "2"
# Kokoro models arrive as tar.bz2 archives, hence tar + bzip2 (pure-Rust
# decoder) for unpacking them.
tar = "0.4"
bzip2 = "0.6"
# Chatterbox: driven through ONNX Runtime directly, since sherpa-onnx has no
# such model, with its GPT-2 tokenizer read by the tokenizers crate - the
# pure-Rust regex backend, so no C library rides along. Behind `chatterbox`,
# which the desktop app turns on: a gigabyte of model is not for a phone.
ort = { workspace = true, optional = true }
# Where ONNX Runtime is made ready, the one place for the cutout's models
# and these: on an Intel Mac it is loaded at first use, and not at all on a
# macOS older than it. concat-vision's manifest also picks the runtime API
# version ort asks for, per target.
concat-vision = { workspace = true, optional = true, features = ["infer"] }
tokenizers = { version = "0.23", optional = true, default-features = false, features = ["fancy-regex"] }
# What a run says about itself: which model loaded, how long a read took,
# where a loop ran out. Through the facade, so the window's log file and a
# terminal both get it; see concat-host's logs.rs.
log = { version = "0.4", features = ["std"] }
# Transcription: whisper.cpp in-process. Built from source by cmake at build
# time; on macOS with Metal so the encoder runs on the GPU. `raw-api` for
# ggml's backend registry: where whisper.cpp is a shared build, its CPU
# kernels are loaded by hand (transcribe.rs, `cpu_backend`).
whisper-rs = { version = "0.16", features = ["raw-api"] }
# Text to speech: the official k2-fsa bindings. Linked statically on the
# desktop Unixes, so the binary carries no extra dylibs.
[target.'cfg(not(any(windows, target_os = "android", target_os = "ios")))'.dependencies]
sherpa-onnx = "1.13"
# Elsewhere it is the shared library.
#
# On Windows it must be: the only static archives k2-fsa publishes are /MT
# (static CRT), and MSVC refuses to link those into a binary whose every
# other object - Rust's own runtime, Skia, whisper - is /MD. LNK2038, on
# every object pair. The /MT code stays sealed inside the DLLs, which ship
# beside the binary.
#
# Windows builds must also set SHERPA_ONNX_LIB_DIR to a directory holding
# only the import .lib files (CI does; ~/.cargo/config.toml locally). Left
# to its own download, the sys crate finds the DLLs next to the import
# libraries and tries to copy them into the "profile directory" - which it
# locates by searching for an ancestor named after $PROFILE, a walk that
# panics under the custom `app` profile. Keeping the DLLs out of the lib
# directory skips that code path; packaging ships them explicitly.
#
# On the phones it is what k2-fsa publishes: one archive of .so files per
# Android ABI, which the app's jniLibs carry, and an xcframework for iOS
# with onnxruntime folded into the one dylib the app embeds. The sys crate
# downloads the right one for the target on its own.
[target.'cfg(any(windows, target_os = "android", target_os = "ios"))'.dependencies]
sherpa-onnx = { version = "1.13", default-features = false, features = ["shared"] }
[target.'cfg(target_os = "macos")'.dependencies]
whisper-rs = { version = "0.16", features = ["metal"] }
# The CoreML provider, so Chatterbox can run on the Mac's own GPU and Neural
# Engine when Settings › Speech asks for it. Pulls pyke's CoreML build of the
# runtime in place of the plain one.
ort = { workspace = true, optional = true, features = ["coreml"] }
# DirectML, so Chatterbox runs on any Direct3D 12 GPU where the runtime
# has the provider. The shipped Windows bundle links Microsoft's CPU build
# instead, so there it runs on the CPU; concat-vision's manifest says why.
[target.'cfg(target_os = "windows")'.dependencies]
ort = { workspace = true, optional = true, features = ["directml"] }
[features]
chatterbox = ["dep:ort", "dep:tokenizers", "dep:concat-vision"]
[lints]
workspace = true