Speech to text

A video you add needs a transcript, and the transcript is usually what you do not have. Speech to text makes one on the computer Parseh runs on: it listens to the video’s sound with Whisper, a speech-recognition model, and hands the page you are adding the video on a timed transcript that you can read, correct and add like any other.

It is optional in every sense. Until you press a button on its page nothing is fetched, Parseh’s own installation is exactly the size it was, and the page a video is added on is the page it always was, with a transcript field you fill in yourself. It is the same on a computer with no graphics card at all.

1. Getting there#

It is not a part of the reading-help page. That page is about reading what nobody has glossed; this is a tool of the page a video is added on, and it has a door of its own. The reading-help page carries one line pointing here.

On a phone the page opens with the phone’s own bar, like its sibling Settings pages, and it does everything the computer’s does.

2. Who may use it#

Any device that has been let in may get, stop and remove anything on this page: the computer, a phone on the Wi-Fi, another computer. The door’s pill says any device let in, and there is no lock line and no dead button anywhere on it.

It is safe to leave open because nothing a device sends becomes anything that is fetched. The only bytes that can ever arrive are the ones this version of Parseh names: the packages of one hash-pinned list, each checked against a SHA-256 Parseh ships with, and complete model packages at fixed revisions, each checked against their own. A model comes from the shared catalogue and a way of running it is one of three. Whoever presses the button gets the same files. Removing a part is refused while an install or a transcription is using it — and a transcription of a YouTube video is using it from the moment it starts, for as long as the video plays, and not only once all its sound has arrived.

3. What is on the page#

The page has these sections, one under the other.

Whisper program comes first, because every model runs on it. It is a row like a model’s: what it costs, where it is kept, Get it and Remove. Pressing Get it on a model brings it too (below), so it is only pressed here to get the program alone.

Automatically run a second Whisper pass is a checkbox. It starts checked and rechecks the first pass’s suspect words after dictionary checks, adding possible alternatives without replacing the transcript or its timestamps. Turn it off to review sooner. The same tool remains available later in Transcribe & review, for a selected section or an individual word.

Whisper models has a Language filter and Installed only checkbox. Standard models stay available for all Whisper languages. The language-specific section lists additional choices for the language you pick. Choose an installed Preferred model to remember it for new videos in that language; Add a video also lets you choose another and remembers the choice on this host.

Each model uses the same installation controls as the reading help: a state with a sign and a word, what it holds or would cost, a bar with the time left while it comes, Stop, and whose work it is, with its licence, on the last line.

Language-specific choices include Persian Fast and Accuracy, slower, Arabic dialectal, Italian, Japanese, French, German, Turkish, English, Hindi Fast and Accuracy candidate, Spanish and Mandarin Chinese. These labels describe intended trade-offs; they do not guarantee fewer errors. Open Model details and limitations for source/package versions and domain or resource limits. A checkpoint awaiting a complete verified package stays visible as Not available, with a reason. An installable package is not automatically a promise of accurate transcription or timing on your recordings.

Choose Get it to download a model. Persian fast, Persian accuracy/slower, Arabic dialectal, Italian, Hindi fast and Spanish are hosted separately in the Parseh organisation on Hugging Face. Other choices use their pinned upstream packages. Parseh checks every file before installation; Stop cancels the download and Carry on resumes it. Models are optional downloads and are not included in the Parseh release. Once installed, every model works offline.

These prepared packages already contain CTranslate2 weights and matching tokenizer, audio-preprocessing and licence assets. Italian and Hindi adapters have already been merged. You do not need PyTorch, PEFT, conversion commands or a model-hosting application to install a Whisper model in Parseh.

Processor says what the transcript will be made on — below. Hardware remains a separate choice from model and language.

Getting a Whisper model gets its speech program. Pressing Get it on a model when the program is not there installs the program first and then the model, under one bar, and the row says so before it starts (this includes the speech program, which comes first). Nothing is fetched by pressing anything else: opening Add a video never installs anything.

Languages lists every language Parseh has, each in its own script and the right-to-left ones right-to-left, with a tick for those Whisper knows. All eleven of Parseh’s are ticked: Persian, Arabic, Italian, Japanese, French, German, Turkish, English, Hindi, Spanish and Chinese. A language you added yourself is ticked if Whisper has a code for it, and marked not offered if it does not.

Under each offered language is an Exact word times row. Its separate, optional CTC network is about 340–361 MiB installed, and its row says the size before Get it, download progress, Stop or Remove, its licence and its attribution. It is not needed to transcribe: Whisper works with no such network. When installed, it lets the add page make word boundaries from the recording while leaving Whisper’s captions unchanged. All eleven offered languages have one; each download is pinned to an immutable public parseh/aligner-<language> revision and hash-checked before it is installed under stt/aligners/<language>/.

A row’s states#

A model that is here while the program is not says so under it — it cannot be used until the program is back.

4. The processor#

CPU · ready. Speech to text will work on this computer. A compatible NVIDIA graphics card can make it considerably faster, but one is not required. That is the first sentence about the processor, and it is true of every computer the program installs on: the CPU is the way it always works, and it is never described as second-class. It computes in int8, on as many threads as the computer has physical cores, at most eight — more is not faster (on the computer it was measured on, 8 threads took 24 seconds where 24 threads took 34).

The graphics card is used only when Parseh has proved it can be, and never on the strength of a name or of an NVIDIA program being installed. The proof is a look, made by a program of its own that never loads a model, that asks three things separately: whether the driver sees a card, whether the speech program’s own build lists a half-precision type the card can compute in, and whether cuBLAS for CUDA 12 loads — counting the card is not enough, because with cuBLAS missing the count is still one and the model even loads, and it is the first sentence of the transcript that fails. The page opens with a look, and Check again looks once more; that is its own button, for after you have installed something the card needed.

It answers one of four ways:

Technical details, folded under the row, say what the look found: the version of the program, the CPU’s compute type and threads, the card, its driver, the compute type chosen for it, where cuBLAS was found, and when the look was made.

Automatic, CPU, NVIDIA GPU#

These are chosen when a video is added, and the page says what each does:

Automatic — recommended.

Uses the card when it is ready and the CPU otherwise, and silently and reliably the CPU when there is no card: Automatic never causes a failed attempt. If the card cannot start the model after all (its memory, most likely), Parseh continues on the CPU and says so.

CPU.

Always uses the CPU, whatever the computer has.

NVIDIA GPU.

Offered only where the card is ready. Chosen and then failing, it says why and does not quietly switch.

The compute type is never asked for. On a card Parseh uses int8_float16 where the card can, then float16; a card that can do neither is found, not ready and the CPU is used.

A card is not always faster. These are figures from one computer, for one 57-second clip, and only that: the turbo model took 24 seconds on the CPU (8 threads), 11 seconds on a GTX 1650 in int8_float16, and 42 seconds on the same card in float16, which is why the card’s type is chosen for it. large-v3 took 109 seconds on the CPU and needed about 6 GB of memory there; on the 4 GB card it did not fit in float16 and ran, in int8_float16, in 48 seconds. Real speech, a real video and a different computer will differ.

How to enable GPU acceleration#

Graphics acceleration needs, for the version of the speech program this Parseh installs (CTranslate2 4.8.2):

• an NVIDIA graphics driver that runs CUDA 12 programs, from the card’s maker or your system’s driver tool;
• cuBLAS for CUDA 12 (libcublas.so.12 on Linux, cublas64_12.dll on Windows), which comes with NVIDIA’s CUDA Toolkit 12: choose a 12.x release from NVIDIA’s archive, because the newest CUDA has a cuBLAS of another number, which this version cannot use.

cuDNN is not needed. The documentation of faster-whisper still says it is; the CTranslate2 the program carries stopped needing it in its 4.6.3. This list is made from one table in Parseh (GPU_NEEDS in lib/getstt.py), next to the version of the program, so a later version that asks for something else changes what this page says with it.

Parseh installs none of it. It never fetches the graphics card’s software, and never changes a driver: those are the computer’s. When you have installed them, press Check again on the page — nothing else is needed, and CPU transcription worked all the time.

Only the builds for Linux and Windows on an x86-64 processor carry card support; on a Mac, or on Linux on ARM, the CPU is what there is.

5. Which computers can have it#

The program is built for Python 3.12 and for five kinds of computer: Linux on x86-64 (glibc 2.28 or newer) and on ARM64, Windows on x86-64, a Mac with Apple silicon (macOS 14 or newer) and an Intel Mac (macOS 13 or newer). On a computer that is none of these — another Python, another processor, an older Mac, a Linux without glibc, no pip — the rows say Not available, with why, and offer no button. On Windows, the program’s files have names up to 125 characters long, so a Parseh folder nested deeper than about a hundred characters is refused with a sentence, unless Windows has long paths turned on.

Another Python, or a Python with no pip.

Parseh’s own environment has Python 3.12 and pip: make it once by double-clicking install.bat on Windows or Parseh.command on a Mac (or running install.sh on Linux), then start Parseh again.

A folder too deep on Windows.

Either move Parseh’s folder nearer the top of a drive, or turn long paths on: open the Registry Editor, go to HKEY_LOCAL_MACHINE\SYSTEM\CurrentControlSet\Control\FileSystem, set the value LongPathsEnabled to 1 (or turn on Enable Win32 long paths in the Group Policy Editor), and restart Windows. Microsoft’s page on it is Maximum Path Length Limitation.

6. Where the files live#

None of it is in Parseh’s own environment: the program is never on the server’s own path (it brings its own numpy), and everything that uses it — the transcription, and the look at the graphics card — is a separate process that ends when Parseh does. stt/ is kept, like dict/ and mt/, by an update: it is in neither release’s list of files, so no update, forward or back, touches a byte of it; git ignores it; and no backup holds it — fetch it again here.

Going back to a0.4.0 leaves stt/ exactly as it is. That version has no page to remove it, and does not read it; if you do not want the gigabytes while you are on it, take the stt/ folder off by hand, and get speech to text again on the version that has the page.

7. Whose work it is#

Each row says whose work it is and its licence, linked, and the licences page lists them all.

• The program: faster-whisper, CTranslate2 and onnxruntime are under the MIT licence; PyAV is under BSD-3-Clause, and the FFmpeg libraries inside its packages report LGPL-3.0-or-later, while the codec libraries beside them keep their own licences; the rest of what the list names, and the backends CTranslate2’s packages bundle (Intel MKL and oneDNN in the Linux one), are under the licence each comes with. Parseh, which is GPL-3.0-or-later, ships none of it: your own press of Get it fetches it from PyPI.
• Standard models: OpenAI’s Whisper large-v3 and large-v3-turbo, converted to CTranslate2’s format (by Systran, and by Mobius Labs for the turbo one; the repository is now dropbox-dash/faster-whisper-large-v3-turbo), under the MIT licence, fetched from Hugging Face at a fixed version.
• Language-specific models retain their source authors, checkpoints and licences in the catalogue, installation rows and complete packages. The converted package’s identity is recorded separately from its source checkpoint.
• Exact-word-time networks: Parseh’s aligner-zh, aligner-ja, aligner-hi, aligner-ar, aligner-fa, aligner-tr, aligner-es, aligner-de, aligner-fr, aligner-it and aligner-en repositories on Hugging Face, fetched at the pinned revision. They run through onnxruntime; their notices and licences travel with each installed row (Apache-2.0, except Hindi MIT and Turkish CC-BY-4.0).

The hosts it talks to, and only when you press a button, are pypi.org and files.pythonhosted.org for the program, and huggingface.co (which answers with a redirect to its own file host) for the models. Once installed, transcribing a film on this computer needs no network at all.