ASR-Kit

Modular Python library for Automatic Speech Recognition. Pass WAV(s) to a unified Transcriber; get TranscriptionResult(s) back. Backed by pluggable model drivers.

Supported: Qwen3-ASR | Cohere-Transcribe | NVIDIA Parakeet | OpenAI Whisper

Setup

# Use one isolated environment per backend model.
# Create one env per model key (e.g. asr-cohere, asr-qwen, asr-whisper).

MODEL_KEY=cohere
conda create -n asr-${MODEL_KEY} python=3.12
conda activate asr-${MODEL_KEY}

# NVIDIA GPU: install a CUDA-enabled PyTorch build first.
# Use a runtime build published by PyTorch (for example cu124),
# not necessarily your exact driver version.
# Recommended (pip wheel):
# pip install torch --index-url https://download.pytorch.org/whl/cu124

pip install "asr-kit[${MODEL_KEY}] @ git+https://github.com/dqvid3/asr-kit.git"

# Optional for Qwen on Ampere+ GPUs:
pip install -U flash-attn --no-build-isolation

Available model keys right now:

• cohere
• parakeet
• qwen
• whisper

Notes:

• The same install pattern works for new backends as they are added: set MODEL_KEY to the backend name.
• Keep backend models in separate environments; their dependency stacks can conflict.
• Use ISO language codes at the ASR-Kit API boundary, such as it, en, or fr. Drivers translate that to the backend-specific format when needed.

Usage

from asr_kit import Transcriber

# Initialize. Use device="mps" on Apple Silicon, or "cuda" on NVIDIA GPUs.
t = Transcriber(model="cohere", device="mps", dtype="bfloat16")

# Single file → TranscriptionResult
result = t.transcribe("audio.wav", language="en")
print(result.text)

# Multiple files → list[TranscriptionResult] with progress spinner
# Cohere uses ISO codes (e.g. "en", "it", "fr") and supports a punctuation toggle
results = t.transcribe(["a.wav", "b.wav"], language="en", punctuation=True, max_new_tokens=256)

# Qwen supports contextual biasing and timestamps when initialized with an aligner.
q = Transcriber(model="qwen", device="cuda", max_new_tokens=256)
results = q.transcribe(
    "audio.wav",
    language="it",
    context="acronyms, product names, speaker names",
    return_timestamps=True,
)

# Parakeet auto-detects language.
p = Transcriber(model="parakeet", device="cuda")
result = p.transcribe("audio.wav", return_timestamps=True)
print(result.text)
print(result.language)

# Whisper supports language detection, translation, prompts, and word timestamps.
w = Transcriber(model="whisper", model_id="turbo", device="auto")
result = w.transcribe("audio.wav", language="it", return_timestamps=True)

# Disable progress spinner
results = t.transcribe("audio.wav", language="en", show_progress=False)

Gated Models & Authentication

Some models (like Cohere-Transcribe) are gated on Hugging Face. To use them:

1. Request access on the CohereLabs/cohere-transcribe-03-2026 page.
2. Authenticate your local environment:
   • CLI: Run huggingface-cli login and enter your token.
   • Code: Pass your token directly: Transcriber(model="cohere", token="hf_...")
   • Env: Set the HF_TOKEN environment variable in your .env file or shell.

Output: TranscriptionResult

field	type	description
text	str	Full transcript
language	str | None	Detected language code
timestamps	list[WordTimestamp] | None	Word-level timing (if requested)
audio_path	str	Absolute source WAV path
model	str	Model identifier

Timestamp Support

return_timestamps=True is checked before transcription:

• qwen: requires Transcriber(model="qwen", use_forced_aligner=True)
• parakeet: supported by the default v3 model
• cohere: not supported
• whisper: supported by all model variants

Adding a Driver

1. Create src/asr_kit/drivers/my_driver.py implementing BaseDriver.
2. Implement load_model, batch_size, and transcribe.
3. Register the key in _DRIVER_REGISTRY in transcriber.py.
4. Add optional deps in pyproject.toml.

License

MIT