Parakeet MLX

An implementation of the Parakeet models - Nvidia's ASR(Automatic Speech Recognition) models - for Apple Silicon using MLX.

Installation

Using uv - recommended way:

uv add parakeet-mlx -U

Or, for the CLI:

uv tool install parakeet-mlx -U

Using pip:

pip install parakeet-mlx -U

CLI Quick Start

parakeet-mlx <audio_files> [OPTIONS]

Arguments

• audio_files: One or more audio files to transcribe (WAV, MP3, etc.)

Options

• --model (default: mlx-community/parakeet-tdt-0.6b-v3, env: PARAKEET_MODEL)

  • Hugging Face repository of the model to use
  • https://huggingface.co/collections/mlx-community/parakeet

• --output-dir (default: current directory)

  • Directory to save transcription outputs

• --output-format (default: srt, env: PARAKEET_OUTPUT_FORMAT)

  • Output format (txt/srt/vtt/json/all)

• --output-template (default: {filename}, env: PARAKEET_OUTPUT_TEMPLATE)

  • Template for output filenames, {parent}, {filename}, {index}, {date} is supported.

• --highlight-words (default: False)

  • Enable word-level timestamps in SRT/VTT outputs

• --verbose / -v (default: False)

  • Print detailed progress information

• --decoding (default: greedy, env: PARAKEET_DECODING)

  • Decoding method to use (greedy or beam)
  • beam is only available at TDT/RNNT models for now

• --chunk-duration (default: 120 seconds, env: PARAKEET_CHUNK_DURATION)

  • Chunking duration in seconds for long audio, 0 to disable chunking

• --overlap-duration (default: 15 seconds, env: PARAKEET_OVERLAP_DURATION)

  • Overlap duration in seconds if using chunking

• --beam-size (default: 5, env: PARAKEET_BEAM_SIZE)

  • Beam size (only used while beam decoding)

• --length-penalty (default: 0.013, env: PARAKEET_LENGTH_PENALTY)

  • Length penalty in beam. 0.0 to disable (only used while beam decoding)

• --patience (default: 3.5, env: PARAKEET_PATIENCE)

  • Patience in beam. 1.0 to disable (only used while beam decoding)

• --duration-reward (default: 0.67, env: PARAKEET_DURATION_REWARD)

  • From 0.0 to 1.0, < 0.5 to favor token logprobs more, > 0.5 to favor duration logprobs more. (only used while beam decoding in TDT models)

• --max-words (default: None, env: PARAKEET_MAX_WORDS)

  • Max words per sentence

• --silence-gap (default: None, env: PARAKEET_SILENCE_GAP)

  • Split sentence if it exceeds silence gap provided (seconds)

• --max-duration (default: None, env: PARAKEET_MAX_DURATION)

  • Max sentence duration (seconds)

• --fp32 / --bf16 (default: bf16, env: PARAKEET_FP32 - boolean)

  • Determine the precision to use

• --full-attention / --local-attention (default: full-attention, env: PARAKEET_LOCAL_ATTENTION - boolean)

  • Use full attention or local attention (Local attention reduces intermediate memory usage)
  • Expected usage case is for long audio transcribing without chunking

• --local-attention-context-size (default: 256, env: PARAKEET_LOCAL_ATTENTION_CTX)

  • Local attention context size(window) in frames of Parakeet model

• --cache-dir (default: None, env: PARAKEET_CACHE_DIR)

  • Directory for HuggingFace model cache. If not specified, uses HF's default cache location (~/.cache/huggingface or values you set in HF_HOME or HF_HUB_CACHE which is essentially $HF_HOME/hub)

Examples

# Basic transcription
parakeet-mlx audio.mp3

# Multiple files with word-level timestamps of VTT subtitle
parakeet-mlx *.mp3 --output-format vtt --highlight-words

# Generate all output formats
parakeet-mlx audio.mp3 --output-format all

Python API Quick Start

Transcribe a file:

from parakeet_mlx import from_pretrained

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")

result = model.transcribe("audio_file.wav")

print(result.text)

Check timestamps:

from parakeet_mlx import from_pretrained

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")

result = model.transcribe("audio_file.wav")

print(result.sentences)
# [AlignedSentence(text="Hello World.", start=1.01, end=2.04, duration=1.03, tokens=[...])]

Do chunking:

from parakeet_mlx import from_pretrained

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")

result = model.transcribe("audio_file.wav", chunk_duration=60 * 2.0, overlap_duration=15.0)

print(result.sentences)

Do beam decoding:

from parakeet_mlx import from_pretrained, DecodingConfig, Beam

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")

config = DecodingConfig(
    decoding = Beam(
        beam_size=5, length_penalty=0.013, patience=3.5, duration_reward=0.67
        # Refer to CLI options for each parameters 
    )
)

result = model.transcribe("audio_file.wav", decoding_config=config)

print(result.sentences)

Use local attention:

from parakeet_mlx import from_pretrained

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")

model.encoder.set_attention_model(
    "rel_pos_local_attn", # Follows NeMo's naming convention
    (256, 256),
)

result = model.transcribe("audio_file.wav")

print(result.sentences)

Specify the sentence split options:

from parakeet_mlx import from_pretrained, DecodingConfig, SentenceConfig

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")

config = DecodingConfig(
    sentence = SentenceConfig(
        # Refer to CLI Options to see what those options does
        max_words=30, silence_gap=5.0, max_duration=40.0
    )
)

result = model.transcribe("audio_file.wav", decoding_config=config)

print(result.sentences)

Boost the word "trunks":

from parakeet_mlx import from_pretrained, DecodingConfig, Beam, boost

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")

config = DecodingConfig(
    decoding=Beam(
        booster=boost.lift(
            boost.prebind(str.lower, 
                # By binding str.lower before passing it to words, you make it case invariant!
                # Of course, if you do this, the words must be in lowercase form else it wouldn't boost anything.
                boost.words({"trunks": 4.0}, boundary=True)
                # Could also 
                # boost.prebind(str.upper, boost.words({"TRUNKS": 4.0}, boundary=True))
            ),
            model.vocabulary,
        )
    )
)

result = model.transcribe("meinthezoo.wav", decoding_config=config)

# before: ...really, really, really long um fronts. And that's that's cool...
# after: ...really, really, really long um trunks. And that's that's cool...
print(result.text)

To get proper casing in some cases (see this issue):

from parakeet_mlx import from_pretrained, DecodingConfig, Beam, boost

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")

# Please note that booster state do NOT carry over each chunks
config = DecodingConfig(
    decoding = Beam(
        booster = boost.lift( # must `lift`!
            boost.first_uppercase(0.5), # log-probs!
        )
    )
)

result = model.transcribe("audio_file.wav", decoding_config=config)

print(result.sentences)

from_pretrained

Using from_pretrained downloads a model from Hugging Face and stores the downloaded model in HF's cache folder. You can specify the cache folder by passing it cache_dir args. It can return one of those Parakeet variants such as: ParakeetTDT, ParakeetRNNT, ParakeetCTC, or ParakeetTDTCTC. For general use cases, the BaseParakeet abstraction often suffices. However, if you want to call variant-specific functions like .decode() and want linters not to complain, typing.cast can be used.

Timestamp Result

• AlignedResult: Top-level result containing the full text and sentences
  • text: Full transcribed text
  • sentences: List of AlignedSentence
• AlignedSentence: Sentence-level alignments with start/end times
  • text: Sentence text
  • start: Start time in seconds
  • end: End time in seconds
  • duration: Between start and end.
  • tokens: List of AlignedToken
• AlignedToken: Word/token-level alignments with precise timestamps
  • text: Token text
  • start: Start time in seconds
  • end: End time in seconds
  • duration: Between start and end.

Boosting

For boosting, these are the primitives that can be composed:

# combine multiple boosting (and others)
def combine(*machines: Machine[Any, A]) -> Machine[tuple, A]: pass 

# scale those particular group of boosting with a factor
def scale(factor: float, machine: Machine[S, A]) -> Machine[S, A]: pass 

# a map! 1-to-1
def premap(f: Callable[[B], A], machine: Machine[S, A]) -> Machine[S, B]: pass 

# an input filter, like python's filter.
def prefilter(keep: Callable[[A], bool], machine: Machine[S, A]) -> Machine[S, A]: pass

# basically 1-to-many expansion. 
# I recommend using this over premap if you are `str.lower`ing, 
# since it handles the case where it outputs multiple letters
def prebind(f: Callable[[B], Iterable[A]], machine: Machine[S, A]) -> Machine[S, B]: pass 

# boost/penalty words weight in logprob, either freely or with the word boundary 
def words(weights: dict[str, float], *, boundary: bool = True) -> Machine[..., str]: pass 

# boost the probability of the first alphanumeric character being uppercase by given weight
def first_uppercase(weight: float) -> Machine[..., str]: pass 

And you always need to call:

lift(machine, model.vocabulary)

before passing to the Beam(booster=...), to make it operate on token-level instead of characters.

Like this:

from parakeet_mlx import from_pretrained, DecodingConfig, Beam, boost

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")

config = DecodingConfig(
    decoding=Beam(
        booster=boost.lift(
            boost.combine(
                boost.first_uppercase(0.5), # encourage first letter being uppercase AND
                boost.scale( # 2X the boosting here
                    2.0,
                    boost.prebind( # everything inside sees lowercased 
                        str.lower,
                        boost.prefilter( # everything inside now ignore dots
                            lambda char: char != ".",
                            boost.premap( # now they think - are space
                                lambda char: " " if char == "-" else char,
                                boost.words({"new york": 2.0, "usa": 1.5}),
                                # so finally, boost New-York, New York, U.S.A, USA
                            ),
                        ),
                    ),
                ),
            ),
            model.vocabulary,
        )
    )
)

result = model.transcribe("audio.wav", decoding_config=config)

print(result.sentences)

For advanced users, the whole boosting system is basically just a weighted Mealy machine with potential-based reward (or weight) shaping and telescoping sum. So you can also express FSMs with it and, hence, express all regular languages if you want. Take a closer look at boost.py if you need.

Streaming Transcription

For real-time transcription, use the transcribe_stream method which creates a streaming context:

from parakeet_mlx import from_pretrained
from parakeet_mlx.audio import load_audio
import numpy as np

model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")

# Create a streaming context
with model.transcribe_stream(
    context_size=(256, 256),  # (left_context, right_context) frames
) as transcriber:
    # Simulate real-time audio chunks
    audio_data = load_audio("audio_file.wav", model.preprocessor_config.sample_rate)
    chunk_size = model.preprocessor_config.sample_rate  # 1 second chunks

    for i in range(0, len(audio_data), chunk_size):
        chunk = audio_data[i:i+chunk_size]
        transcriber.add_audio(chunk)

        # Access current transcription
        result = transcriber.result
        print(f"Current text: {result.text}")

        # Access finalized and draft tokens
        # transcriber.finalized_tokens
        # transcriber.draft_tokens

Streaming Parameters

• context_size: Tuple of (left_context, right_context) for attention windows

  • Controls how many frames the model looks at before and after current position
  • Default: (256, 256)

• depth: Number of encoder layers that preserve exact computation across chunks

  • Controls how many layers maintain exact equivalence with non-streaming forward pass
  • depth=1: Only first encoder layer matches non-streaming computation exactly
  • depth=2: First two layers match exactly, and so on
  • depth=N (total layers): Full equivalence to non-streaming forward pass
  • Higher depth means more computational consistency with non-streaming mode
  • Default: 1

• keep_original_attention: Whether to keep original attention mechanism

  • False: Switches to local attention for streaming (recommended)
  • True: Keeps original attention (less suitable for streaming)
  • Default: False

Low-Level API

To transcribe log-mel spectrum directly, you can do the following:

import mlx.core as mx
from parakeet_mlx.audio import get_logmel, load_audio
from parakeet_mlx import DecodingConfig

# Load and preprocess audio manually
audio = load_audio("audio.wav", model.preprocessor_config.sample_rate)
mel = get_logmel(audio, model.preprocessor_config)

# Generate transcription with alignments
# Accepts both [batch, sequence, feat] and [sequence, feat]
# `alignments` is list of AlignedResult. (no matter if you fed batch dimension or not!)
alignments = model.generate(mel, decoding_config=DecodingConfig())

Todo

• Add CLI for better usability
• Add support for other Parakeet variants
• Streaming input (real-time transcription with transcribe_stream)
• Option to enhance chosen words' accuracy

Acknowledgments

• Thanks to Nvidia for training these awesome models and writing cool papers and providing nice implementation.
• Thanks to MLX project for providing the framework that made this implementation possible.
• Thanks to audiofile and audresample, numpy, librosa for audio processing.
• Thanks to dacite for config management.

License

Apache 2.0