scriptrs

scriptrs is fast speech-to-text with NVIDIA's Parakeet TDT 0.6B v2 and v3 models. Inference runs in Rust on Apple Core ML or NVIDIA CUDA, with no PyTorch or ONNX Runtime.

pip install scriptrs

The package needs Python 3.11 or newer. One wheel covers every Python version.

Use

import scriptrs

pipeline = scriptrs.Pipeline.from_pretrained("parakeet/v2")  # downloads the model once
result = pipeline.transcribe_file("meeting.mp3")

print(result.text)
for chunk in result.chunks:
    print(f"{chunk.start:7.2f} {chunk.end:7.2f} {chunk.text}")

Choosing a model

Models are named "<family>/<version>". Both Parakeet models come with the same package, and you pick one when you load the pipeline:

english = scriptrs.Pipeline.from_pretrained("parakeet/v2")
multilingual = scriptrs.Pipeline.from_pretrained("parakeet/v3")

scriptrs.Model.PARAKEET_V2 and scriptrs.Model.PARAKEET_V3 work too, and so do NVIDIA's model ids, such as "nvidia/parakeet-tdt-0.6b-v3". Each model downloads separately the first time you load it, so loading only the one you use downloads only that one.

Parakeet v3 detects the language of each window by itself. To fix it, pass an ISO 639-1 code:

result = multilingual.transcribe_file("interview.wav", language="de")
print(scriptrs.Model.PARAKEET_V3.languages)  # ['bg', 'hr', 'cs', ...]

Audio

transcribe_file reads WAV, FLAC, MP3, Ogg Vorbis, and AAC in MP4/M4A, and converts it to 16 kHz mono. Audio already in memory has to be 16 kHz mono float32, such as a NumPy array:

result = pipeline.transcribe(samples, language="en")

Recordings of any length work. Audio longer than one model window is split into overlapping windows and the transcripts are merged. For recordings that are mostly silence or noise, load the pipeline with vad=True to split at detected speech instead.

To transcribe many clips, such as utterances, use transcribe_many. It prepares the next clips while the model runs and, on CUDA, batches the clips' windows together, up to max_batch in one pass (32 by default, at most 64):

results = pipeline.transcribe_many([clip_a, clip_b, clip_c])

Each result has the full text, timed chunks, and every token in tokens with its start, end, and confidence.

Command line

scriptrs transcribe meeting.mp3
scriptrs transcribe interview.wav --model parakeet/v3 --language de --json

With one file, the transcript is printed alone. With several, each one follows a header, as head prints them:

$ scriptrs transcribe a.wav b.wav
==> a.wav <==
First transcript.

==> b.wav <==
Second transcript.

--json prints one JSON object per line, with the file, text, timed chunks, and tokens.

To install only the command line, in its own environment, use uv:

uv tool install scriptrs

CUDA

The Linux and Windows wheels need an NVIDIA driver, but no CUDA toolkit: they install cuBLAS from the nvidia-cublas-cu12 package. The default precision="fp16" is the fastest. "tf32" keeps the weights in FP32, and "fp32" matches NVIDIA NeMo's FP32 output token for token:

pipeline = scriptrs.Pipeline.from_pretrained("parakeet/v3", precision="fp32")

The Windows wheel builds and passes its tests in CI on a runner without a GPU, but nobody has transcribed with it on a Windows GPU yet. It loads the driver's nvcuda.dll and the nvidia-cublas-cu12 package's cublas64_12.dll and cublasLt64_12.dll. If transcription fails there, please open an issue.

Models

from_pretrained() downloads the models from Hugging Face into cache_dir= when given, else the Hugging Face hub cache that HF_HUB_CACHE or HF_HOME sets, else ~/.cache/huggingface/hub. On macOS a model is about 0.6 GB of Core ML models from avencera/scriptrs-models. On Linux and Windows it is about 2.5 GB of weights from avencera/parakeet-tdt-0.6b-v2-cuda or avencera/parakeet-tdt-0.6b-v3-cuda.

To use models you manage yourself, load them with scriptrs.Pipeline.from_dir(path, "parakeet/v2"), using the layout in the scriptrs README.

Errors

Invalid arguments, such as an unknown model name or samples that are not 1-D float32, raise ValueError or TypeError. Everything else that fails raises scriptrs.ScriptrsError. Its code is a stable value, such as unsupported_language, model_download, audio_decode, or cuda, so you can branch on it without parsing the message.