fscript 1.0.3

Fast local transcription for large lectures with NVIDIA Parakeet ONNX

fast-transcript

fast-transcript is a local lecture transcription CLI built to beat the usual Apple Silicon tradeoff: either fast but flaky, or accurate but painfully slow.

On the development machine, this project handled 30 minutes in 2!* while staying around 2.51 GB RSS on the long run. In the same local test set, it beat mlx-whisper, insanely-fast-whisper, and parakeet-mlx.

_{* Benchmark run on a MacBook Pro M1. The exact long-run measurement was 29m47s of Portuguese lecture audio transcribed in about 2m14s (13.38x real-time).}

The CLI binary is called fscript:

fscript lecture.mp3
fscript lecture.mp3 notes/
fscript lecture.mp3 --text
fscript lecture.mp3 --text=plain
fscript lecture.mp3 --raw
fscript lecture.mp3 --srt
fscript lecture.mp3 --vtt
fscript lecture.mp3 --json
fscript lecture.mp3 --diarize lseend-dihard3
fscript lecture.mp3 -D --json --raw
fscript lecture.mp3 -n 2

--srt and --vtt subtitle output are experimental.

That is the whole point of this project. One command. Large audio. No babysitting.

Why this exists

I wanted a tool for transcribing long classes and lectures quickly on a laptop while still using the computer for normal work.

The existing options I tested had clear problems for this use case:

• insanely-fast-whisper was far too slow on this Mac once it fell back to CPU
• mlx-whisper was solid, but slower than I wanted for long lecture workflows
• parakeet-mlx had excellent memory numbers, but drifted into English on longer Portuguese segments unless heavily tuned

fast-transcript packages the ONNX Parakeet path that held up best in practice.

What it does

• downloads the default Parakeet TDT 0.6B v3 int8 model automatically if it is missing
• stores the extracted model in a persistent per-user application data directory
• keeps the downloaded tarball in the user cache directory
• accepts local audio/video files in formats supported by ffmpeg
• accepts remote http(s) video/audio URLs supported by yt-dlp
• prefers platform-provided manual subtitles for remote URLs when available
• falls back to downloading remote audio and transcribing locally when only auto-captions exist or no captions exist
• auto-converts unsupported audio to 16 kHz mono PCM16 WAV
• uses 120s chunks with 2s overlap by default
• runs local speaker diarization by default via fluidaudiocli process --mode offline
• writes <input>.speakers.txt next to the input unless you choose a different output path
• can alternatively write raw transcript text via --text, with timestamps on by default and --text=plain as the opt-out
• can alternatively write experimental subtitle files via --srt or --vtt
• can alternatively write speaker-aware text via --speakers, defaulting to HH:MM:SS - SPEAKER_01: ...
• cleans pathological repeated-word runs such as we we we we into we... we by default, with --raw as the opt-out
• stays quiet by default: concise progress in the terminal, transcript JSON on disk
• shows a spinner and chunk progress bar on interactive terminals

Install

Requirements

• ffmpeg
• ffprobe
• yt-dlp for remote URLs, or uvx yt-dlp
• fluidaudiocli on PATH if you want speaker-aware diarization
  • when it is missing, fscript warns and continues without diarization
  • use -d, --diarize, or --diarize lseend-dihard3 to require diarization explicitly

Install with Homebrew

brew tap brenorb/tap
brew install fast-transcript

On Apple Silicon macOS, the tap also installs fluidaudio-cli, so the default speaker-aware mode works out of the box.

If you prefer the fully-qualified formula name:

brew install brenorb/tap/fast-transcript

If you previously used the old tap name, migrate with:

brew uninstall brenorb/fast-transcript/fast-transcript
brew untap brenorb/fast-transcript
brew tap brenorb/tap
brew install fast-transcript

On Apple Silicon macOS, Homebrew now installs fast-transcript from a proper bottle. On Linux x86_64, the formula still installs from the published release binary.

PyPI / uv

The PyPI package name for this project is fscript so the target UX is:

uvx fscript lecture.mp3
uv tool install fscript

The repo already includes platform wheel builds for:

• macOS arm64
• Linux x86_64

PyPI publishing is currently enabled for:

• macOS arm64

See docs/pypi-publishing.md for the release workflow details.

Install a prebuilt binary directly

Download the archive for your platform from the GitHub Releases page, then put fscript on your PATH.

Build from source

cargo install --git https://github.com/brenorb/fast-transcript

Or from a local clone:

cargo install --path .

On macOS, the build now auto-detects the active Xcode or Command Line Tools Clang runtime directories so cargo test keeps linking even if your Rust toolchain points at a stale libclang_rt.osx path.

Development

For local development, install the checked-in git hooks once per clone:

scripts/install-git-hooks.sh

The pre-commit hook blocks Rust commits that would fail cargo fmt --check in CI.

Quick start

fscript lecture.mp3
fscript https://www.youtube.com/watch?v=QSdh8Gj0mEg

This will:

1. ensure the default model exists
2. normalize the audio if needed
3. transcribe with the default chunking strategy
4. diarize with the default coreml backend
5. write lecture.speakers.txt
6. print the final absolute transcript path to stdout

For remote URLs, the default speaker-aware flow is:

1. inspect the URL with yt-dlp
2. download the remote audio
3. run the normal local transcription + diarization pipeline

If you switch to -D or --no-diarization, fscript can still use platform-provided manual subtitles directly when they are available unless you also force --local.

Usage

fscript <media-or-url> [output-path]
fscript <media-or-url> -o output-path
fscript <media-or-url> --stdout
fscript <media-or-url> -
fscript <media-or-url> --speakers
fscript <media-or-url> --speakers=plain
fscript <media-or-url> --speakers=timestamps
fscript <media-or-url> --text
fscript <media-or-url> --text=plain
fscript <media-or-url> --raw
fscript <media-or-url> --json
fscript <media-or-url> --srt
fscript <media-or-url> --vtt
fscript <media-or-url> --diarize lseend-dihard3
fscript <media-or-url> -D --json --raw
fscript <media-or-url> -n 2
fscript --version

<media-or-url> can be a local audio file, a local video file, or a supported remote media URL.

When fscript writes the transcript to a file, it keeps progress and human-readable status on stderr and prints only the final absolute transcript path on stdout. That makes it easy to compose in shell scripts:

out=$(fscript lecture.mp3)
open "$out"

If the explicit output-path already exists as a directory, fscript writes the default filename for the chosen mode inside that directory.

Optional overrides:

fscript lecture.wav custom-output.txt
fscript lecture.wav exports/
fscript lecture.wav -o custom-output.txt
fscript lecture.wav --stdout
fscript lecture.wav --speakers
fscript lecture.wav --speakers=plain
fscript lecture.wav --speakers=timestamps
fscript lecture.wav --text
fscript lecture.wav --text=plain
fscript lecture.wav --raw
fscript lecture.wav --json
fscript lecture.wav --srt
fscript lecture.wav --vtt
fscript lecture.wav --diarize lseend-dihard3
fscript lecture.wav -D --json --raw
fscript lecture.wav -n 2
fscript lecture.wav --chunk 180 --overlap 3
fscript lecture.wav --chunk 0
fscript lecture.wav --model-dir ./models/parakeet/custom-copy
fscript lecture.wav --model-package ./models/parakeet-v3-int8.tar.gz
fscript lecture.wav --model-url https://example.com/parakeet-v3-int8.tar.gz
fscript https://www.youtube.com/watch?v=QSdh8Gj0mEg
fscript https://www.youtube.com/watch?v=QSdh8Gj0mEg --local

Raw text output modes:

• --text: transcript text with segment timestamps, one line per segment with HH:MM:SS - ...
• --text=plain: transcript text without timestamps or speaker labels
• when --text is active and you do not pass an explicit output path, the default file becomes <input>.transcript.txt

Cleaning mode:

• cleaning is on by default and affects only the output being written for that invocation
• --raw: disables output cleaning for that invocation
• it applies to JSON, speakers, text, SRT, and VTT outputs
• it is intentionally conservative and leaves ordinary repetition alone

Subtitle output modes:

• --srt: experimental SubRip subtitle file
• --vtt: experimental WebVTT subtitle file
• subtitle output is still experimental and may change
• if diarization is active, subtitle cues include normalized speaker labels such as SPEAKER_01: ...
• when --srt is active and you do not pass an explicit output path, the default file becomes <input>.srt
• when --vtt is active and you do not pass an explicit output path, the default file becomes <input>.vtt

Speaker-aware output modes:

• --speakers: speaker-aware output with timestamps, for example 00:12:34 - SPEAKER_01: ...
• --speakers=timestamps: explicit alias for the default speaker-aware timestamped output
• --speakers=plain: speaker-aware output without timestamps, for example SPEAKER_01: ...
• if diarization is disabled or a segment has no speaker label, the line falls back to plain segment text without an UNKNOWN: prefix
• when no output mode is passed, --speakers is the default
• when --speakers is active and you do not pass an explicit output path, the default file becomes <input>.speakers.txt

Environment overrides:

• FSCRIPT_MODEL_DIR
• FSCRIPT_MODEL_PACKAGE
• FSCRIPT_MODEL_URL
• FSCRIPT_DIARIZATION_BINARY

Optional diarization

fscript automatically enables speaker diarization when fluidaudiocli is available. If the helper is missing and you did not request diarization explicitly, fscript falls back to plain transcript segments after printing a warning on stderr.

By default, it:

1. runs the normal Parakeet ASR flow first
2. releases the ASR model
3. runs a separate fluidaudiocli diarization subprocess
4. merges diarization windows into ASR segments by temporal overlap

Modes:

• -d / --diarize: enable diarization explicitly with the default coreml model
• -d coreml / --diarize coreml: FluidInference/speaker-diarization-coreml path via fluidaudiocli process --mode offline
• -d lseend-dihard3 / --diarize lseend-dihard3: alternate FluidInference/ls-eend-coreml DIHARD III path via fluidaudiocli lseend --variant dihard3
  • defaults to --threshold 0.3
• -D / --no-diarization: skip diarization entirely
• --backend=coreml|lseend-dihard3|none: legacy alias kept for backwards compatibility

Controls:

• -n N / --num-speakers N is forwarded only to the default coreml backend
• -t N / --threshold N overrides the default diarization threshold for lseend-dihard3
• lseend-dihard3 does not support --num-speakers; use the default threshold or override it with -t / --threshold

If you explicitly request -d, --diarize, or a concrete diarization model and fluidaudiocli is missing, fscript returns a clear backend error instead of silently falling back.

Defaults

• model dir:
  • macOS: ~/Library/Application Support/fast-transcript/models/parakeet-tdt-0.6b-v3-int8
  • Linux: ~/.local/share/fast-transcript/models/parakeet-tdt-0.6b-v3-int8
• model package cache:
  • macOS: ~/Library/Caches/fast-transcript/parakeet-v3-int8.tar.gz
  • Linux: ~/.cache/fast-transcript/parakeet-v3-int8.tar.gz
• model URL: https://huggingface.co/brenorb/parakeet-tdt-0.6b-v3-int8-onnx-bundle/resolve/main/parakeet-v3-int8.tar.gz?download=1
• chunk seconds: 120
• chunk overlap seconds: 2
• default diarization mode: coreml when fluidaudiocli is available
• cleaning: on
• default output path: <input>.speakers.txt
• output path with --json: <input>.transcript.json
• output path with --text: <input>.transcript.txt
• output path with --srt: <input>.srt
• output path with --vtt: <input>.vtt
• output path with --speakers: <input>.speakers.txt

Benchmarks

These are local development benchmarks, not universal claims. They were run on the same Apple Silicon Mac used during development, using a Portuguese lecture clip and the same broader workflow comparison.

2-minute lecture clip

Engine	Setup	Speed	Peak RSS	Notes
fast-transcript	Parakeet ONNX	13.06x real-time	2.25 GB	Best balance of speed and reliability
mlx-whisper	whisper-large-v3-turbo	5.25x	1.70 GB	Good quality, slower
parakeet-mlx	tuned for quality	4.92x	1.29 GB	Needed substantial tuning
parakeet-mlx	raw greedy	10.16x	0.57 GB	Faster on short audio, drifted into English on longer PT-BR
insanely-fast-whisper	whisper-large-v3 CPU	0.30x	6.18 GB	Accurate, but too slow here
insanely-fast-whisper	MPS + fallback	0.31x	3.04 GB	Small gain, same general problem

Long lecture run

Engine	Audio	Speed	Peak RSS	Notes
fast-transcript	29m47s lecture	13.38x real-time	2.51 GB	Stable long run with default chunking

Practical reading

• fast-transcript was not the absolute fastest thing we saw in every synthetic case
• it was the best result once long Portuguese lecture audio, transcript quality, and unattended runs all mattered at the same time
• that is the target workload for this repo

Output format

Default output is speaker-aware text and includes:

• segment timestamps
• speaker labels when diarization returns them
• cleaned repeated-word runs unless you pass --raw

JSON output via --json includes:

• merged transcript text
• model path
• original input path
• prepared WAV path
• whether a remote URL used manual subtitles or the local model
• whether ffmpeg normalization was used
• load time
• transcribe time
• chunk configuration
• per-chunk timing
• transcript segments
• optional speaker_diarization metadata

When diarization is enabled, each transcript segment may include:

• speaker

Alternative output modes:

• --speakers: speaker-aware text with timestamps
• --speakers=timestamps: explicit speaker-aware text with timestamps
• --speakers=plain: speaker-aware text without timestamps
• --text: transcript text with segment timestamps
• --text=plain: transcript text without timestamps
• --json: structured JSON benchmark/transcript payload
• --srt: experimental subtitle file
• --vtt: experimental subtitle file

Motivation

This project is optimized for large lectures and classes, including files in the 30-minute to 2-hour range, where:

• startup friction matters
• background CPU usage matters
• memory spikes matter
• brittle hand-tuned command lines become a tax

The design goal is not “highest benchmark on a cherry-picked GPU server”. The goal is “transcribe big local lecture audio fast enough that you actually keep using it”.

Inspiration

This project was heavily informed by:

• Handy
• GLaDOS
• transcribe-rs

In particular, the ONNX Parakeet path here was shaped by the packaging and implementation ideas used in Handy and GLaDOS.

Default model bundle

The default auto-download bundle is published in our own Hugging Face model repository:

• brenorb/parakeet-tdt-0.6b-v3-int8-onnx-bundle

This keeps the default install path tied to the exact validated tarball instead of an app-specific blob host.

License

MIT