Implementation:Elevenlabs Elevenlabs python ForcedAlignmentWordResponseModel

Attribute	Value
Sources	src/elevenlabs/types/forced_alignment_word_response_model.py
Domains	Forced Alignment, Audio Timing, Transcription
Last Updated	2026-02-15

Overview

Description

The ForcedAlignmentWordResponseModel represents a single word with its timing information and confidence score from the ElevenLabs forced aligner. In addition to the word text and its start/end times in seconds, this model includes a loss field representing the average alignment loss/confidence score for the word, calculated from its constituent characters. Lower loss values generally indicate higher confidence in the alignment accuracy.

Usage

This model is returned as part of forced alignment results, representing word-level timing data. It is used for tasks such as word-level subtitle synchronization, audio editing at word boundaries, and quality assessment of alignment results (via the loss field). Each instance represents one word's temporal position and alignment confidence in the audio.

Code Reference

Source Location

src/elevenlabs/types/forced_alignment_word_response_model.py

Class Signature

class ForcedAlignmentWordResponseModel(UncheckedBaseModel):
    """
    Model representing a single word with its timing information from the aligner.
    """
    ...

Import Statement

from elevenlabs.types import ForcedAlignmentWordResponseModel

Base Class

UncheckedBaseModel (from elevenlabs.core.unchecked_base_model)

I/O Contract

Field	Type	Required	Description
text	str	Yes	The word that was transcribed.
start	float	Yes	The start time of the word in seconds.
end	float	Yes	The end time of the word in seconds.
loss	float	Yes	The average alignment loss/confidence score for this word, calculated from its constituent characters.

Usage Examples

from elevenlabs import ElevenLabs

client = ElevenLabs(api_key="your_api_key")

# Obtain forced alignment results (from a dubbing or alignment endpoint)
alignment_result = client.audio_native.get_alignment(
    audio_id="audio_abc123"
)

# Iterate through word-level alignment data
for word in alignment_result.words:
    duration = word.end - word.start
    print(
        f"Word '{word.text}': "
        f"{word.start:.3f}s - {word.end:.3f}s "
        f"(duration: {duration:.3f}s, loss: {word.loss:.4f})"
    )

# Filter for words with high alignment confidence (low loss)
high_confidence_words = [
    w for w in alignment_result.words if w.loss < 0.1
]
print(f"High confidence words: {len(high_confidence_words)}")

# Generate SRT-style subtitles from word timing
for i, word in enumerate(alignment_result.words):
    start_h = int(word.start // 3600)
    start_m = int((word.start % 3600) // 60)
    start_s = word.start % 60
    print(f"{start_h:02d}:{start_m:02d}:{start_s:06.3f} -> {word.text}")

Related Pages

• ForcedAlignmentCharacterResponseModel - Character-level forced alignment timing
• Alignment - Chunk-level audio-text alignment data