parakeet.cpp

Fast speech recognition with NVIDIA's Parakeet models in pure C++.

Built on axiom — a lightweight tensor library with automatic Metal GPU acceleration. No ONNX runtime, no Python runtime, no heavyweight dependencies. Just C++ and one tensor library that outruns PyTorch MPS.

~27ms encoder inference on Apple Silicon GPU for 10s audio (110M model) — 96x faster than CPU. FP16 support for ~2x memory reduction.

Supported Models

Model	Class	Size	Type	Description
tdt-ctc-110m	ParakeetTDTCTC	110M	Offline	English, dual CTC/TDT decoder heads
tdt-600m	ParakeetTDT	600M	Offline	Multilingual, TDT decoder
eou-120m	ParakeetEOU	120M	Streaming	English, RNNT with end-of-utterance detection
nemotron-600m	ParakeetNemotron	600M	Streaming	Multilingual, configurable latency (80ms–1120ms)
sortformer	Sortformer	117M	Streaming	Speaker diarization (up to 4 speakers)
diarized	DiarizedTranscriber	110M+117M	Offline	ASR + diarization → speaker-attributed words

All ASR models share the same audio pipeline: 16kHz mono WAV → 80-bin Mel spectrogram → FastConformer encoder.

Quick Start

cppCopy
#include <parakeet/parakeet.hpp>

parakeet::Transcriber t("model.safetensors", "vocab.txt");
t.to_gpu();   // optional — Metal acceleration
t.to_half();  // optional — FP16 inference (~2x memory reduction)

auto result = t.transcribe("audio.wav");
std::cout << result.text << std::endl;

Features

• Multiple decoders — CTC greedy, TDT greedy, CTC beam search, TDT beam search (switch at call site)
• Word timestamps — Per-word start/end times and confidence scores on all decoders
• Beam search + LM — CTC and TDT beam search with optional ARPA n-gram language model fusion
• Phrase boosting — Context biasing via token-level trie for domain-specific vocabulary
• Batch transcription — Multiple files in one batched encoder forward pass
• VAD preprocessing — Silero VAD strips silence before ASR; timestamps auto-remapped
• GPU acceleration — Metal via axiom's MPSGraph compiler (96x speedup on Apple Silicon)
• FP16 inference — Half-precision weights and compute (~2x memory reduction)
• Streaming — EOU and Nemotron models with chunked audio input
• Speaker diarization — Sortformer (up to 4 speakers), combinable with ASR for speaker-attributed words
• C API — Flat extern "C" FFI for Python, Swift, Go, Rust, and other languages
• Multi-format audio — WAV, FLAC, MP3, OGG with automatic resampling

See examples/ for code demonstrating each feature.

Install

Prebuilt binaries are attached to each GitHub release for macOS arm64, macOS x86_64, and Linux x86_64. Download the tarball for your platform and extract:

bashCopy
tar -xzf parakeet-v0.1.0-macos-arm64.tar.gz
cd parakeet-v0.1.0-macos-arm64
# On macOS, clear the Gatekeeper quarantine attribute first:
xattr -dr com.apple.quarantine .
./bin/parakeet --help

The archive ships a self-contained bin/parakeet (and bin/example-server) plus lib/libaxiom with @rpath/$ORIGIN set so the binaries resolve their dependencies relative to the install dir — drop the directory anywhere. The C-API headers under include/parakeet/ are included for embedders.

Build from source

bashCopy
git clone --recursive https://github.com/frikallo/parakeet.cpp
cd parakeet.cpp
make build
make test

Requirements: C++20 (Clang 14+ or GCC 12+), CMake 3.20+, macOS 13+ for Metal GPU.

macOS: building requires the full Xcode install (not just Command Line Tools), because axiom compiles its Metal shaders with xcrun metal and xcrun metallib — those ship only with Xcode. If you just want to run parakeet, use the prebuilt tarball above; the .metallib is embedded into the shipped libaxiom.dylib and runs without any Xcode/CLT install on the user side.

Convert Weights

bashCopy
# Download from HuggingFace
huggingface-cli download nvidia/parakeet-tdt_ctc-110m --include "*.nemo" --local-dir .

# Convert to safetensors
pip install safetensors torch
python scripts/convert_nemo.py parakeet-tdt_ctc-110m.nemo -o model.safetensors

The converter supports all model types: 110m-tdt-ctc (default), 600m-tdt, eou-120m, nemotron-600m, sortformer.

bashCopy
python scripts/convert_nemo.py checkpoint.nemo -o model.safetensors --model 600m-tdt

Silero VAD weights:

bashCopy
python scripts/convert_silero_vad.py -o silero_vad_v5.safetensors

Examples

Example	Description
basic	Simplest transcription (~20 lines)
timestamps	Word/token timestamps with confidence
beam-search	CTC/TDT beam search with optional ARPA LM
phrase-boost	Context biasing for domain vocabulary
batch	Batch transcription of multiple files
vad	Standalone VAD and ASR+VAD preprocessing
gpu	Metal GPU + FP16 with timing comparison
stream	EOU streaming transcription
nemotron	Nemotron streaming with latency modes
diarize	Sortformer speaker diarization
diarized-transcription	ASR + diarization combined
c-api	Pure C99 FFI usage
cli	Full CLI with all options

Using as a Library

CMake find_package

After installing (make install or cmake --install build):

cmakeCopy
find_package(Parakeet REQUIRED)
target_link_libraries(myapp PRIVATE Parakeet::parakeet)

CMake add_subdirectory

cmakeCopy
add_subdirectory(third_party/parakeet.cpp)
target_link_libraries(myapp PRIVATE Parakeet::parakeet)

pkg-config

bashCopy
g++ -std=c++20 myapp.cpp $(pkg-config --cflags --libs parakeet) -o myapp

Architecture

Offline Models

Built on a shared FastConformer encoder (Conv2d 8x subsampling → N Conformer blocks with relative positional attention):

Model	Class	Decoder	Use case
CTC	ParakeetCTC	Greedy argmax or beam search (+LM)	Fast, English-only
RNNT	ParakeetRNNT	Autoregressive LSTM	Streaming capable
TDT	ParakeetTDT	LSTM + duration prediction, greedy or beam search (+LM)	Better accuracy than RNNT
TDT-CTC	ParakeetTDTCTC	Both TDT and CTC heads	Switch decoder at inference

Streaming Models

Built on a cache-aware streaming FastConformer encoder with causal convolutions and bounded-context attention:

Model	Class	Decoder	Use case
EOU	ParakeetEOU	Streaming RNNT	End-of-utterance detection
Nemotron	ParakeetNemotron	Streaming TDT	Configurable latency streaming

Diarization

Model	Class	Architecture	Use case
Sortformer	Sortformer	NEST encoder → Transformer → sigmoid	Speaker diarization (up to 4 speakers)

Benchmarks

Measured on Apple M3 16GB with simulated audio input (Tensor::randn). Times are per-encoder-forward-pass (Sortformer: full forward pass).

Encoder throughput — 10s audio:

Model	Params	CPU (ms)	GPU (ms)	GPU Speedup
110m (TDT-CTC)	110M	2,581	27	96x
tdt-600m	600M	10,779	520	21x
rnnt-600m	600M	10,648	1,468	7x
sortformer	117M	3,195	479	7x

110m GPU scaling across audio lengths:

Audio	CPU (ms)	GPU (ms)	RTF	Throughput
1s	262	24	0.024	41x
5s	1,222	26	0.005	190x
10s	2,581	27	0.003	370x
30s	10,061	32	0.001	935x
60s	26,559	72	0.001	833x

GPU acceleration powered by axiom's Metal graph compiler which fuses the full encoder into optimized MPSGraph operations.

bashCopy
make bench ARGS="--110m=models/model.safetensors --tdt-600m=models/tdt.safetensors"

Roadmap

Tier 1 — High Impact

• Confidence scores — Per-token and per-word confidence from token log-probs
• Phrase boosting — Token-level trie context biasing during decode
• Beam search — CTC prefix beam search and TDT time-synchronous beam search
• N-gram LM fusion — ARPA language models scored at word boundaries

Audio & I/O

• Multi-format audio — WAV, FLAC, MP3, OGG via dr_libs + stb_vorbis
• Automatic resampling — Windowed sinc interpolation (Kaiser, 16-tap)
• Load from memory — read_audio(bytes, len), float/int16 buffers
• Audio duration query — Header-only duration without full decode
• Progress callbacks — Stage reporting for long files
• Streaming from raw PCM — Direct microphone buffer feeding

Tier 2 — Production Readiness

• Diarized transcription — ASR + Sortformer → speaker-attributed words
• VAD — Silero VAD v5, standalone + ASR preprocessing
• Batch inference — Padded multi-file encoder forward pass
• Long-form chunking — Overlapping windows for audio >30s
• Neural LM rescoring — N-best reranking with Transformer LM

Tier 3 — Ecosystem

• C API — Flat C interface for FFI from any language
• FP16 inference — Half-precision weights and compute
• Model quantization — INT8/INT4 for mobile deployment
• Hotword detection — Trigger phrase detection
• Speaker embeddings — Speaker verification from Sortformer/TitaNet

Notes

• Audio: 16kHz mono (WAV, FLAC, MP3, OGG — auto-detected and resampled)
• Offline models have ~4-5 minute audio length limits; use streaming models for longer audio
• GPU acceleration requires Apple Silicon with Metal support

License

MIT