SentenceConfig is a dataclass that controls how transcribed text is split into sentences based on punctuation, silence gaps, word count, or duration.
Class Definition
from parakeet_mlx import SentenceConfig
config = SentenceConfig(
max_words=None,
silence_gap=None,
max_duration=None
)
Fields
Maximum number of words allowed in a single sentence. When the next token would exceed this limit, a sentence break is created.Set to None to disable word-based splitting.Example: max_words=30
silence_gap
float | None
default:"None"
Minimum silence duration (in seconds) that triggers a sentence split. When the gap between tokens exceeds this threshold, a new sentence is created.Set to None to disable silence-based splitting.Example: silence_gap=5.0
max_duration
float | None
default:"None"
Maximum duration (in seconds) allowed for a single sentence. When a sentence reaches this duration, it is split even if no other conditions are met.Set to None to disable duration-based splitting.Example: max_duration=40.0
Splitting Behavior
Sentences are automatically split at punctuation marks (., !, ?, 。, ？, ！). Additional splits occur when any of the following conditions are met:
- Punctuation: Token contains sentence-ending punctuation
- Word limit: Next token would exceed
max_words (if set)
- Silence gap: Gap between current and next token exceeds
silence_gap (if set)
- Duration limit: Sentence duration exceeds
max_duration (if set)
Examples
Default Configuration
By default, sentences are only split at punctuation:
from parakeet_mlx import from_pretrained, DecodingConfig, SentenceConfig
model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")
# Default: only split on punctuation
config = DecodingConfig(sentence=SentenceConfig())
result = model.transcribe("audio.wav", decoding_config=config)
for sentence in result.sentences:
print(sentence.text)
Limit Words per Sentence
Split long sentences to keep them under 30 words:
from parakeet_mlx import from_pretrained, DecodingConfig, SentenceConfig
model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")
config = DecodingConfig(
sentence=SentenceConfig(max_words=30)
)
result = model.transcribe("audio.wav", decoding_config=config)
for sentence in result.sentences:
word_count = len([t for t in sentence.tokens if " " in t.text])
print(f"{word_count} words: {sentence.text}")
Split on Long Silences
Create sentence breaks when silence exceeds 5 seconds:
from parakeet_mlx import from_pretrained, DecodingConfig, SentenceConfig
model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")
config = DecodingConfig(
sentence=SentenceConfig(silence_gap=5.0)
)
result = model.transcribe("audio.wav", decoding_config=config)
for sentence in result.sentences:
print(f"[{sentence.start:.1f}s - {sentence.end:.1f}s] {sentence.text}")
Limit Sentence Duration
Keep sentences under 40 seconds:
from parakeet_mlx import from_pretrained, DecodingConfig, SentenceConfig
model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")
config = DecodingConfig(
sentence=SentenceConfig(max_duration=40.0)
)
result = model.transcribe("audio.wav", decoding_config=config)
for sentence in result.sentences:
print(f"Duration: {sentence.duration:.1f}s - {sentence.text}")
Combine Multiple Constraints
Use all splitting criteria together:
from parakeet_mlx import from_pretrained, DecodingConfig, SentenceConfig
model = from_pretrained("mlx-community/parakeet-tdt-0.6b-v3")
config = DecodingConfig(
sentence=SentenceConfig(
max_words=30, # Max 30 words
silence_gap=5.0, # Split on 5+ second silences
max_duration=40.0 # Max 40 second duration
)
)
result = model.transcribe("audio.wav", decoding_config=config)
for i, sentence in enumerate(result.sentences, 1):
print(f"Sentence {i}: [{sentence.start:.1f}s - {sentence.end:.1f}s]")
print(f" {sentence.text}")
print()
Use Cases
Subtitle Generation
For video subtitles, limit words and duration to fit on screen:
config = DecodingConfig(
sentence=SentenceConfig(
max_words=15, # Short lines for readability
max_duration=5.0 # Quick screen changes
)
)
Meeting Transcription
For meeting notes, split on natural pauses:
config = DecodingConfig(
sentence=SentenceConfig(
silence_gap=3.0, # Split on speaker pauses
max_duration=30.0 # Keep paragraphs manageable
)
)
Podcast Transcription
For long-form content with natural flow:
config = DecodingConfig(
sentence=SentenceConfig(
max_words=40, # Allow longer sentences
silence_gap=8.0 # Only split on significant pauses
)
)