Skip to main content
DecodingConfig is a dataclass that controls how the model decodes audio and splits text into sentences.

​
Class Definition

from parakeet_mlx import DecodingConfig, Greedy, Beam, SentenceConfig

config = DecodingConfig(
    decoding=Greedy(),
    sentence=SentenceConfig()
)

​
Fields

​
decoding
Union[Greedy, Beam]
default:"Greedy()"
The decoding strategy to use. Can be either Greedy() for greedy decoding or Beam() for beam search decoding.See Greedy and Beam below for configuration options.
​
sentence
SentenceConfig
default:"SentenceConfig()"
Configuration for how to split transcribed text into sentences.See SentenceConfig for configuration options.

​
Greedy

Greedy decoding selects the most likely token at each step. This is the fastest decoding method. 
from parakeet_mlx import DecodingConfig, Greedy

config = DecodingConfig(decoding=Greedy())
Greedy decoding has no configuration parameters.

​
Beam

Beam search decoding explores multiple hypotheses to find better transcriptions. Currently only available for TDT models. 
from parakeet_mlx import DecodingConfig, Beam

config = DecodingConfig(
    decoding=Beam(
        beam_size=5,
        length_penalty=1.0,
        patience=1.0,
        duration_reward=0.7
    )
)

​
Fields

​
beam_size
int
default:"5"
Number of hypotheses to explore simultaneously. Larger values may improve accuracy but increase computation time.Example: beam_size=5
​
length_penalty
float
default:"1.0"
Penalty applied based on sequence length. Higher values favor longer sequences.
  • 1.0: No penalty
  • > 1.0: Favor longer sequences
  • < 1.0: Favor shorter sequences
Example: length_penalty=0.013
​
patience
float
default:"1.0"
Controls how many candidate hypotheses to explore. Higher values allow more exploration.The maximum number of candidates is beam_size * patience.Example: patience=3.5
​
duration_reward
float
default:"0.7"
TDT-only parameter. Controls the balance between token logprobs and duration logprobs.
  • 0.0: Only consider token logprobs
  • 0.5: Equal weight to both
  • 1.0: Only consider duration logprobs
Example: duration_reward=0.67

​
Examples

​
Greedy Decoding

from parakeet_mlx import from_pretrained, DecodingConfig, Greedy

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

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

print(result.text)

​
Beam Search 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
    )
)

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

​
With Sentence Configuration

from parakeet_mlx import from_pretrained, DecodingConfig, Beam, SentenceConfig

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

config = DecodingConfig(
    decoding=Beam(beam_size=5),
    sentence=SentenceConfig(
        max_words=30,
        silence_gap=5.0,
        max_duration=40.0
    )
)

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

for sentence in result.sentences:
    print(f"{sentence.start:.2f}s - {sentence.end:.2f}s: {sentence.text}")

Build docs developers (and LLMs) love

Get started for freeTalk to us