antenna-stt 0.4.0

Rust-powered Speech-to-Text toolkit with Whisper, Distil-Whisper, and streaming transcription

Antenna

Rust-powered multi-model Speech-to-Text toolkit for Python.

Supported Models

Model	Backend	GPU	Description
Whisper	Candle	CUDA, Metal	OpenAI's encoder-decoder model (tiny → large-v3)
Distil-Whisper	Candle	CUDA, Metal	2x faster distilled Whisper variants
Wav2Vec2	ONNX	CUDA, TensorRT	Meta's CTC-based model
Parakeet	sherpa-rs	CUDA, DirectML	NVIDIA's FastConformer-TDT (600M params, blazing fast)

Installation

CPU Version

pip install antenna-stt

GPU Version (CUDA)

# Option A: Pre-built wheel
pip install antenna-stt[cuda]

# Option B: Build from source (requires CUDA toolkit)
git clone https://github.com/fiction-ai-studios/antenna.git
cd antenna
pip install maturin
maturin build --release --features cuda
pip install target/wheels/antenna_stt-*.whl

Development Installation

git clone https://github.com/fiction-ai-studios/antenna.git
cd antenna
uv venv && source .venv/bin/activate
uv add --dev maturin pytest pytest-asyncio

# Build options:
uv run maturin develop --release                    # CPU only
uv run maturin develop --release --features cuda    # + Whisper GPU
uv run maturin develop --release --features onnx    # + Wav2Vec2
uv run maturin develop --release --features sherpa  # + Parakeet
uv run maturin develop --release --features sherpa-cuda  # + Parakeet GPU

Quick Start

Basic Transcription

import antenna

# One-liner
result = antenna.transcribe("speech.wav", model_size="base")
print(result.text)

# With more control
audio = antenna.load_audio("speech.wav")
audio = antenna.preprocess_for_whisper(audio)
model = antenna.WhisperModel.from_size("base", device="cuda")
result = model.transcribe(audio)

for segment in result.segments:
    print(f"[{segment.start:.2f}s] {segment.text}")

Unified Model Registry

import antenna

# Load ANY model through unified API
model = antenna.load_model("whisper/base", device="cuda")
model = antenna.load_model("distil-whisper/distil-small.en", device="cpu")
model = antenna.load_model("wav2vec2/base-960h", device="cpu")      # Requires ONNX
model = antenna.load_model("parakeet/tdt-0.6b-v2", device="cuda")   # Requires sherpa

# List all available models
for m in antenna.list_models():
    print(f"{m.id}: {m.description}")

Streaming Transcription

import antenna

# Real-time chunk-by-chunk processing
transcriber = antenna.StreamingTranscriber.from_model_id(
    "whisper/tiny",
    device="cpu",
    config=antenna.StreamingConfig.realtime()
)

for chunk in audio_chunks:
    events = transcriber.process_chunk(chunk)
    for event in events:
        if event.is_final():
            print(event.text())

transcriber.flush()

Async Streaming

import asyncio
import antenna

async def transcribe():
    transcriber = antenna.AsyncStreamingTranscriber.from_model_id(
        "whisper/tiny", device="cuda"
    )
    for chunk in audio_chunks:
        events = await transcriber.process_chunk_async(chunk)
        for event in events:
            if event.is_final():
                print(event.text())
    await transcriber.flush_async()

asyncio.run(transcribe())

Feature Flags

Feature	Description	Models Enabled
cuda	Candle CUDA GPU	Whisper, Distil-Whisper
metal	Candle Metal GPU (macOS)	Whisper, Distil-Whisper
onnx	ONNX Runtime	Wav2Vec2
onnx-cuda	ONNX with CUDA	Wav2Vec2 (GPU)
sherpa	sherpa-rs backend	Parakeet
sherpa-cuda	sherpa with CUDA	Parakeet (GPU)

Build with multiple features:

uv run maturin develop --release --features "cuda,onnx,sherpa-cuda"

Parakeet Models (NEW)

NVIDIA Parakeet is incredibly fast (can transcribe 60min audio in ~1 second).

Setup:

# 1. Build with sherpa feature
uv run maturin develop --release --features sherpa-cuda

# 2. Download model
mkdir -p ~/.cache/antenna/parakeet
wget https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2
tar xvf sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2 -C ~/.cache/antenna/parakeet/

Usage:

import antenna

model = antenna.load_model("parakeet/tdt-0.6b-v2", device="cuda")
result = model.transcribe(audio)
print(result.text)

Variants:

• parakeet/tdt-0.6b-v2 - English (recommended)
• parakeet/tdt-0.6b-v3 - Multilingual (25 languages)

Audio Processing

import antenna

# Load any format (WAV, MP3, FLAC, OGG, M4A)
audio = antenna.load_audio("podcast.mp3")

# Analyze
stats = antenna.analyze_audio(audio)
print(f"RMS: {stats.rms_db:.1f} dB, Peak: {stats.peak_db:.1f} dB")

# Process
audio = antenna.trim_silence(audio, threshold_db=-40)
audio = antenna.normalize_audio(audio, method="rms", target_db=-20)
audio = antenna.preprocess_audio(audio, target_sample_rate=16000, mono=True)

# Save
antenna.save_audio(audio, "processed.wav")

Model Selection Guide

Model	Size	Speed	Quality	Use Case
Whisper tiny	39M	★★★★★	★★	Quick tests
Whisper base	74M	★★★★	★★★	General use
Whisper large-v3	1.5G	★	★★★★★	Best quality
Distil-Whisper	~350M	★★★★	★★★★	Fast + accurate
Wav2Vec2	95-317M	★★★	★★★	CTC decoding
Parakeet	600M	★★★★★	★★★★★	Fastest + accurate

GPU Availability Check

import antenna

print(f"CUDA: {antenna.is_cuda_available()} ({antenna.cuda_device_count()} devices)")
print(f"Metal: {antenna.is_metal_available()}")
print(f"ONNX: {antenna.is_onnx_available()}")
print(f"ONNX CUDA: {antenna.is_onnx_cuda_available()}")

Testing

cargo test --features sherpa    # Rust tests (129)
uv run pytest tests/ -v         # Python tests (226+)

Roadmap

• ✅ Whisper/Distil-Whisper (Candle)
• ✅ Wav2Vec2 (ONNX)
• ✅ Parakeet (sherpa-rs)
• ✅ Streaming API with VAD
• ✅ Async streaming
• ✅ GPU support (CUDA, Metal, DirectML)
• 🔲 Canary (NeMo format)
• 🔲 Conformer (sherpa-rs)
• 🔲 Production HTTP/WebSocket server

License

MIT

Acknowledgments

• PyO3 - Rust-Python bindings
• Candle - ML framework
• sherpa-onnx - Parakeet inference
• ONNX Runtime - Cross-platform inference
• OpenAI Whisper - Original Whisper model