praasper 0.7.3

VAD-Enhanced ASR Framework for Researchers

Praasper

Setup | Usage | Mechanism

Praasper is a speech processing framework designed to help researchers transcribe audio files into word-level timestamps — from a single word to a complete sentence — with high accuracy in both transcription and timestamps.

In Praasper, the pipeline has four stages. First, long recordings are split at natural pauses via pause-aware chunking. Second, VAD (Praditor) performs coarse DBSCAN clustering followed by fine sliding-window boundary detection — automatically calibrated per file via a two-stage grid search (amp × eps_ratio → numValid refinement). Third, ASR (Fun-ASR-Nano) transcribes each VAD-bounded segment with word-level timestamps. Fourth, timestamps are aligned to VAD intervals by temporal overlap and exported as a Praat TextGrid file.

How to use

Here is one of the simplest examples:

import praasper

model = praasper.init_model()
model.annote("data_folder")

Here are the parameters you can pass to init_model and annote:

Param	Default	Description
infer_mode	"local"	ASR backend: "local" for on-device FunASR-Nano, "api" for DashScope cloud API.
device	"auto"	Hardware for local inference: "auto", "cuda", or "cpu". Ignored in API mode.
ASR	FunAudioLLM/Fun-ASR-Nano-2512	Advanced: override the default local ASR model. See FunASR model zoo.
api_key	None	DashScope API key. Required when infer_mode="api". Can also be set via DASHSCOPE_API_KEY env var.
cache_dir	None	Directory for caching ASR models. When set, HF_HOME / MODELSCOPE_CACHE / TRANSFORMERS_CACHE are redirected here.
effort	"normal"	Grid search level: "low" (3 combos), "normal" (22 combos), or "high" (100 combos). "high" adds cutoff0 sweep and finer amp steps. Can be set on init_model and overridden per-run via annote().
input_path	—	Path to the folder where audio files are stored.
seg_dur	15.	Segment large audio into pieces, in seconds.
min_pause	0.2	Minimum pause duration between two utterances, in seconds.
skip_existing	False	Skip files that already have an output .TextGrid.
verbose	False	Print verbose progress messages during processing.

Here are code examples showing how to use these parameters:

import praasper

# Local inference (default)
model = praasper.init_model()

# Local inference on GPU
model = praasper.init_model(device="cuda")

# DashScope cloud API
model = praasper.init_model(infer_mode="api", api_key="sk-...")

# Custom local ASR model (advanced)
model = praasper.init_model(ASR="iic/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch")

# Low-effort grid search (3 combos, fastest)
model = praasper.init_model(effort="low")

# High-effort grid search (100 combos, most thorough VAD calibration)
model = praasper.init_model(effort="high")

model.annote(
    input_path="data_folder",
    min_pause=.8,
    seg_dur=15.
)

Custom VAD Parameters

By default, Praasper automatically calibrates VAD parameters for each recording via a grid search — no manual tuning is required. For advanced users who need custom parameters, a dict or .txt file can be passed directly.

Parameters use the internal Praditor format: a dict with onset and offset sections, each containing amp, cutoff0, cutoff1, numValid, and eps_ratio. Praasper keeps onset and offset identical internally.

Here is a code example showing how to override the default parameters for a specific audio file. The VAD parameters are passed as a dict with onset and offset sections:

import praasper

model = praasper.init_model()

# Define custom VAD parameters
custom_params = {
    "onset":  {"amp": "1.05", "cutoff0": "60", "cutoff1": "10800", "numValid": "475", "eps_ratio": "0.05"},
    "offset": {"amp": "1.05", "cutoff0": "60", "cutoff1": "10800", "numValid": "475", "eps_ratio": "0.05"},
}

model.annote(
    input_path="data_folder",
    params=custom_params,
)

Alternatively, save the parameters to a .txt file and pass the file path instead:

model.annote(
    input_path="data_folder",
    params="/path/to/custom_params.txt",
)

In both cases, Praasper will use your custom VAD parameters instead of running the auto grid search. To export the current parameters to a .txt file for later reuse:

model.export_params("/path/to/custom_params.txt")

ASR: local vs. cloud

Praasper supports two ASR backends, chosen via the infer_mode parameter:

infer_mode	Backend	Best for
"local" (default)	FunASR-Nano — lightweight, runs on laptop CPU/GPU	Offline work, no API costs, privacy
"api"	DashScope (AliCloud) — cloud ASR with stronger accuracy	High-accuracy needs, server deployments

Local mode (infer_mode="local")

model = praasper.init_model()                    # auto-detect GPU/CPU
model = praasper.init_model(device="cuda")       # force GPU
model = praasper.init_model(device="cpu")        # force CPU

The default model is FunAudioLLM/Fun-ASR-Nano-2512, which supports Chinese, English, and Japanese with word-level timestamps. Power users can swap in a different FunASR model via the ASR parameter.

API mode (infer_mode="api")

model = praasper.init_model(infer_mode="api", api_key="sk-...")
# or set DASHSCOPE_API_KEY environment variable:
# model = praasper.init_model(infer_mode="api")

Note: DashScope requires audio files to be hosted at a public URL (HTTP/HTTPS/OSS). Local file paths are not supported in API mode.

Mechanism

Praasper processes audio in four stages:

1. Pause-aware chunking. Long recordings are split into segments (default 15 s) at natural pauses detected by the VAD, placing boundaries at silence-gap midpoints. If no gap is found, the threshold relaxes until a boundary can be placed. This preserves utterance integrity across chunks.

2. Voice Activity Detection (VAD). Praasper uses Praditor, a DBSCAN-based detector. The first stage clusters the amplitude envelope to separate speech from silence into broad candidate segments. The second stage applies a sliding-window detector with locally estimated noise thresholds to place onset and offset boundaries at millisecond precision. By default, Praasper auto-calibrates per recording via effort:

Effort	Stage 1	Stage 2	Total combos
"low"	amp=[1.2] × eps_ratio=[0.02,0.03,0.04]	skipped	3
"normal"	amp=[1.1,1.2,1.3] × eps_ratio=[0.02–0.05, 6 steps]	numValid=[500–5000]	22
"high"	8 amp × 6 eps_ratio × cutoff0=[0,200]	numValid=[500–5000]	100

Stage 1 maximises onset boundary SNR; Stage 2 refines numValid (DBSCAN min points) around the winner. Manual tuning is available but not required.

3. Automatic Speech Recognition (ASR). Each VAD-bounded segment is transcribed by Fun-ASR-Nano, a lightweight model producing word-level timestamps. It supports Chinese (Mandarin and 7 dialects: Wu, Cantonese, Min, Hakka, Gan, Xiang, Jin), English, and Japanese. For higher accuracy, DashScope cloud ASR is available via infer_mode="api".

4. Overlap matching and export. ASR word timestamps are matched to VAD intervals by maximum temporal overlap. Unmatched words are assigned to the nearest interval by distance. Words in the same interval are concatenated; empty intervals are discarded. Adjacent overlapping intervals are merged. Processing runs in parallel (4 workers). Output is a Praat-compatible TextGrid file.

Advanced users can override the auto-calibration by supplying custom VAD parameters via a Python dict or .txt file (see Custom VAD Parameters).

Setup

pip installation

pip install -U praasper

If you have a successful installation and don't care about GPU acceleration, you can stop right here.

GPU Acceleration (Windows/Linux)

Currently, Praasper utilizes Fun-ASR-Nano from FunASR as the default local ASR engine. Cloud ASR is also available via DashScope (infer_mode="api").

FunASR automatically detects the best currently available device to use. But you still need to first install the GPU-support version of torch in order to enable CUDA acceleration.

• For macOS users, only CPU is supported as the processing device.
• For Windows/Linux users, the priority order should be: CUDA -> CPU.

If you have no experience in installing CUDA, follow the steps below:

First, go to command line and check the latest CUDA version your system supports:

nvidia-smi

Results should pop up like this (It means that this device supports CUDA up to version 12.9).

| NVIDIA-SMI 576.80                 Driver Version: 576.80         CUDA Version: 12.9     |

Next, go to NVIDIA CUDA Toolkit and download the latest version, or whichever version that fits your system/need.

Lastly, install torch that fits your CUDA version. Find the correct pip command in this link.

Here is an example for CUDA 12.9:

pip install --force-reinstall torch torchaudio --index-url https://download.pytorch.org/whl/cu129

After installation, verify that PyTorch can see your GPU:

python -c "import torch; print('CUDA available:', torch.cuda.is_available()); print('CUDA version:', torch.version.cuda); print('GPU:', torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'N/A')"

Expected output (example for RTX 3050):

CUDA available: True
CUDA version: 13.0
GPU: NVIDIA GeForce RTX 3050

If CUDA available shows False, double-check that you installed a CUDA-enabled torch wheel (from https://download.pytorch.org/whl/cu*) — not the CPU-only wheel from PyPI. Re-run the pip install --force-reinstall command above with the correct --index-url for your driver.

⚠️ RTX 30-series users: Avoid torch 2.8.0 with cu129 — a known CUDA 12.9 GPU dispatch bug causes 5× slower model loading and garbled output on Ampere GPUs. Use CUDA 12.6+ or 13.0+ instead (e.g., --index-url https://download.pytorch.org/whl/cu130).

You can also verify your CUDA toolkit installation:

nvcc --version

The nvcc version should match or be compatible with the CUDA version reported by nvidia-smi and torch.version.cuda. Note that nvidia-smi shows the maximum supported CUDA version (driver-level), while nvcc --version shows the installed toolkit version — either one works as long as PyTorch's CUDA is compatible with your driver.

(Advanced) uv installation

uv is also highly recommended for a much faster installation. First, make sure uv is installed in your default environment:

pip install uv

Then, create a virtual environment (e.g., .venv):

uv venv .venv

You should see a new .venv folder appear in your project directory. (You may also want to restart the terminal.)

Lastly, install praasper (by prefixing pip with uv):

uv pip install -U praasper

For CUDA support, here is an example for downloading torch that fits CUDA 12.9:

uv pip install --force-reinstall torch torchaudio --index-url https://download.pytorch.org/whl/cu129