native-whisperx
native-whisperx is a Rust-first WhisperX parity workflow repository built
from the published moritzbrantner-* Rust crates.
This repository owns application composition:
- CLI commands
- workflow configuration
- output file writing
- parity checks against Python WhisperX
- setup documentation for local model bundles and Hugging Face cache resolution
It does not own the reusable ASR, transcript, model-runtime, alignment, or
speaker primitives. Those remain in
rust-packages and are
consumed here as published crates.
Crate Layout
crates/native-whisperx # reusable workflow library
crates/native-whisperx-cli # native-whisperx CLI binary
Cargo Install
The first Cargo release keeps the two-crate shape. Publish
native-whisperx first, then publish native-whisperx-cli. The CLI crate is
the Cargo install package, and it installs the native-whisperx terminal
command:
These installed-binary smoke commands are no-resource offline checks. They do not transcribe media, download models, use CUDA, call Python WhisperX, read Hugging Face credentials, or require a local smoke media root.
Installed transcription commands use the same local model bundle, cache, and
resource requirements as repository examples. Install success only proves that
the CLI package and native-whisperx command are available; transcription
readiness still depends on the requested Whisper, alignment, diarization, VAD,
translation, CUDA, Python WhisperX compatibility, and gated Hugging Face
resources. Delegated Feature paths remain delegated until replaced by explicit
Rust-Native Parity work.
Maintainer release gates and the manual library-first, CLI-second publish
checklist are in docs/publish-plan.md.
Workflow
media or samples
-> moritzbrantner-audio-analysis-transcription
-> moritzbrantner-text-transcripts::TranscriptionContract
-> default wav2vec2 alignment unless disabled
-> optional speaker diarization
-> optional segment-level post-ASR translation
-> WhisperX JSON, Native JSON, SRT, WebVTT, or plain text outputs
Feature Flags
| Feature | Purpose |
|---|---|
native |
Native Candle Whisper and wav2vec2 alignment composition. Enabled by default. |
translation |
Helsinki-NLP OPUS-MT/Marian post-ASR segment translation and curated planning/runtime APIs. Enabled by default; translation still runs only when configured. |
cuda |
CUDA-backed Candle execution. Opt in when a local CUDA toolchain is available. |
media-decode |
FFmpeg-backed finite non-WAV media/container decode through the audio I/O crate. Enabled by default. |
diarization |
Heuristic speaker diarization composition. |
onnx-diarization |
Explicit ONNX speaker embedding diarization path. |
pyannote-vad |
Native pyannote ONNX VAD path. Enabled by default so Automatic Workflow Selection can resolve pyannote VAD resources lazily for native --diarize. |
pyannote-diarization |
Native pyannote community diarization bundle path. Enabled by default so Automatic Workflow Selection can resolve pyannote diarization resources lazily for native --diarize. |
whisperx-compat |
External Python WhisperX command compatibility and parity checks. |
Default installed CLI builds include translation plus the pyannote VAD and
pyannote community diarization code paths needed by Automatic Workflow
Selection. They do not bundle or eagerly resolve model files: help, version,
and Speaker Directory commands remain no-resource offline checks, while
translation and diarization resources resolve only when those workflows run.
Default features do not enable CUDA, external Python WhisperX compatibility,
parity resources, live-feed resource checks, ONNX speaker embedding
diarization, or Silero VAD. Explicit --no-default-features library and CLI
builds remain translation-free; enable translation explicitly for a minimal
translation-capable build.
Commands
Import existing WhisperX JSON:
Inspect the request shape for local model bundles:
Run native transcription with an explicit Whisper bundle:
Run native transcription from a locally cached Hugging Face Whisper model:
Run native diarization with Automatic Workflow Selection:
Automatic Workflow Selection is a Workflow Composition concept. For native
finite --diarize, it chooses pyannote VAD plus
pyannote/speaker-diarization-community-1 when the user has not explicitly
chosen lower-level VAD or diarization model settings. It is distinct from the
WhisperX Parity contract and from Rust-Native Parity evidence; it changes how a
native workflow is composed, not transcript output contracts.
Automatic native --diarize looks for pyannote resources in --model-dir,
then standard Hugging Face cache roots. --model-cache-only is a hard
no-download guarantee, so missing automatic VAD or diarization resources fail
before transcription. Without cache-only, the future lookup order allows a
download path, but current pyannote automatic downloads are not wired to a
bundle hydrator yet; missing resources still fail before transcription with
setup guidance. Native automatic selection uses environment or standard Hugging
Face auth state for future/prepared cache workflows, not CLI token strings.
Do not put token values in commands or reports.
Prepare local pyannote resources under --model-dir before using automatic
cache-only diarization. See
docs/model-bundles.md
for accepted local directory and Hugging Face cache layouts plus maintainer
real-resource checks.
Finite Media Inputs
Default native-whisperx and native-whisperx-cli builds include finite media
decode support. WAV files continue to use the existing native WAV reader path.
Non-WAV finite media files route through the FFmpeg-backed media decode path
when the required runtime tools are installed.
The guaranteed finite input set is wav, mp3, m4a, aac, flac, ogg,
opus, mp4, mov, mkv, and webm. Other FFmpeg-decodable files may work
on a best-effort basis, but they are not part of the guaranteed support set.
Video files are transcribed from the selected/default audio track only; video
frames are not analyzed.
Builds using --no-default-features do not implicitly include finite non-WAV
media decode. Enable media-decode explicitly for minimal builds that still
need FFmpeg-backed media/container input support.
Input Pattern Expansion applies to finite media paths. Transcribe multiple
files by passing concrete paths or app-expanded wildcard patterns. Relative and
absolute paths are accepted. Quoted patterns such as 'audio/*.wav' and
'media/*.mp4' are expanded by native-whisperx before transcription, so they do
not depend on shell glob behavior:
When --output-dir is omitted, the default json transcript uses Input-Local
Output and is written beside each input file. When --output-dir is supplied,
all outputs use that shared directory and native-whisperx fails before
transcription if two inputs would write the same output basename. --basename
is rejected with multiple expanded inputs.
Run native post-ASR translation:
This path transcribes source-language segments with native Whisper, translates segment text with the configured OPUS-MT Marian model, and preserves segment timings for downstream writers.
Reusable Translation, Progress, and Cancellation
The library's curated translation registry covers English, German, French,
Spanish, Italian, Portuguese, Dutch, and Polish. TranslationPlan chooses a
deterministic validated direct model when available or records an explicit
two-leg Pivot Translation through English. translate_transcription borrows an
immutable source result and returns a separate translated result with ordered
model provenance. The source text, language, segments, word and character
timings, diagnostics, and metadata remain unchanged and usable after either
success or failure.
TranslationConfig is owned by the library rather than the CLI. Embedding
applications can choose an explicit model_bundle, an application-owned
model_dir, and model_cache_only without constructing CLI argument types.
Cache-only translation never downloads missing assets.
Finite applications observe ordered phase, model resolution/download,
model load/reuse, translation-leg, output, completion, failure, and
cancellation facts through TranscriptionProgressObserver. The cloneable
CancellationHandle cooperatively stops run_with_control and
run_many_with_control at safe Workflow Composition boundaries and returns a
typed cancellation outcome without writing later outputs.
Live Feed Transcription uses a separate LiveTranscriptionProgressObserver so
operational progress is not confused with Live Transcript Events. Passing the
same cancellation handle to
LivePcmIngestionSession::ingest_reader_with_control stops before the next
Near-Live Window, retains stable final events, discards unstable partial text,
and emits LiveSessionEndReason::Cancelled.
Migration note for embedding applications: exhaustive matches over
LiveSessionEndReason must add the new Cancelled arm. Both progress event
enums are #[non_exhaustive], so downstream matches require a wildcard arm.
The library crate README, also rendered on docs.rs, contains the complete API
contract and compatibility guidance.
Run the ignored manual cache-only native ASR smoke when SMOKE_ROOT contains
the required audio and Hugging Face cache layout:
Detailed setup is in docs/model-bundles.md.
Run the ignored real FFmpeg finite media decode smoke when validating local runtime media support. It generates tiny temporary audio and video containers from FFmpeg filter sources, then verifies each non-WAV input reaches native cache-only ASR model resolution after media decode:
RUN_NATIVE_FFMPEG_MEDIA_DECODE_SMOKE=1
This smoke is intentionally opt-in and is not part of default offline CI. It
requires local ffmpeg and ffprobe, but does not require model bundles, CUDA,
Python WhisperX, network access, or committed binary media fixtures.
Run native-vs-Python WhisperX parity:
Preflight local parity resources, generate ignored Python WhisperX 3.8.6 goldens, then run the local ASR parity fixture suite:
--format json writes WhisperX-compatible JSON. Use --format native-json
when you need the Rust transcript contract shape.
Alignment is enabled by default. Use --no-align / --no_align to skip it.
Use --whisper-bundle and --alignment-bundle for explicit local bundles.
Without --whisper-bundle, native ASR can resolve supported Whisper models
through the Hugging Face cache or download path; --model-cache-only requires
the files to already exist locally and never downloads.
--translation-model reuses --model-dir and --model-cache-only for
translation model resolution unless --translation-bundle is supplied.
Published-Crate Requirement
This workspace intentionally uses crates.io dependencies so a clean checkout can
resolve and test without a sibling ../rust-packages repository. The published
dependency closure is tracked in docs/publish-plan.md.
During local co-development, keep local Cargo patches outside commits. One
option is to put overrides in a local Cargo config and keep that config
untracked, for example by adding .cargo/ to .git/info/exclude before
creating .cargo/config.toml:
[]
= { = "../rust-packages/crates/runtime/runtime-core" }
= { = "../rust-packages/crates/audio/audio-analysis-speakers" }
= { = "../rust-packages/crates/audio/audio-analysis-transcription" }
= { = "../rust-packages/crates/runtime/model-runtime" }
= { = "../rust-packages/crates/text/text-model-runtime" }
= { = "../rust-packages/crates/text/text-transcripts" }
= { = "../rust-packages/crates/video/video-analysis-core" }
Add any transitive unpublished crates to the same local override only for local validation. Do not commit patch entries to this repository.
Pull Request CI
The default pull request workflow runs offline Rust gates on GitHub-hosted Ubuntu runners:
The feature-matrix rows are compile-only gates. They cover the external WhisperX compatibility bridge, translation, media decode, heuristic diarization, Silero VAD, ONNX diarization, default pyannote VAD and diarization packaging, and the combined offline optional feature set without running model inference.
These checks do not require local model bundles, Python WhisperX, CUDA devices,
Hugging Face tokens, ONNX Runtime dynamic-library configuration, or self-hosted
parity resources. Real-resource parity checks remain in the opt-in
parity-fixtures workflow.
Runtime-only or intentionally constrained combinations stay outside pull request CI:
| Feature or path | Gate | Reason and expected failure mode |
|---|---|---|
silero-vad runtime transcription |
parity-preflight / parity-fixtures |
Compile is covered in PR CI, but execution requires ORT_DYLIB_PATH and a local Silero ONNX bundle. Missing resources fail preflight before model execution. |
onnx-diarization runtime transcription |
parity-preflight / parity-fixtures |
Compile is covered in PR CI, but execution requires ONNX Runtime plus local diarization model artifacts. Missing resources fail preflight before model execution. |
pyannote-vad and pyannote-diarization full-resource parity |
parity-fixtures final-full-surface |
Meaningful validation requires local pyannote bundles, expected WhisperX goldens, Python WhisperX resources, and gated Hugging Face access for the delegated reference path. Missing resources are reported by preflight. |
automatic native --diarize pyannote cache and download boundary |
manual full-resource commands in docs/model-bundles.md |
Cache-only prepared-cache runs validate automatic pyannote lookup without downloads. The no-cache boundary command documents the current fail-before-transcription behavior until a pyannote download hydrator is wired. |
cuda |
manual/report-only throughput ladder | The CUDA feature requires a compatible local CUDA toolchain and device, so it is not a GitHub-hosted pull request gate. |