Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/docling-project/docling/llms.txt

Use this file to discover all available pages before exploring further.

​
Overview

The Audio backend processes audio files and converts speech to text using Automatic Speech Recognition (ASR) models. It supports various audio formats and uses Whisper-based models for high-quality transcription.

​
Features

  • Multi-format support - MP3, WAV, FLAC, OGG, M4A, WEBM
  • Video audio tracks - Extracts and transcribes audio from video files
  • Whisper models - Multiple model sizes for speed/accuracy tradeoffs
  • Timestamp information - Preserves timing data in transcription
  • Speaker diarization - Optional speaker identification (model-dependent)
  • Language detection - Automatic language identification
  • Multi-language support - Supports 90+ languages

​
Supported Formats

Audio formats

  • MP3 (.mp3)
  • WAV (.wav)
  • FLAC (.flac)
  • OGG (.ogg)
  • M4A (.m4a)
  • OPUS (.opus)

Video formats (audio track)

  • MP4 (.mp4)
  • WEBM (.webm)
  • MKV (.mkv)
  • AVI (.avi)

​
Usage

​
Basic Transcription

from docling.document_converter import DocumentConverter

converter = DocumentConverter()
result = converter.convert("audio.mp3")

doc = result.document

# Access transcription
for item, _ in doc.iterate_items():
    if isinstance(item, TextItem):
        print(item.text)

​
With Pipeline Options

from docling.document_converter import DocumentConverter, AudioFormatOption
from docling.datamodel.pipeline_options import AsrPipelineOptions
from docling.datamodel import asr_model_specs

pipeline_options = AsrPipelineOptions(
    asr_options=asr_model_specs.WHISPER_BASE  # Use base model
)

converter = DocumentConverter(
    format_options={
        AudioFormatOption: AudioFormatOption(
            pipeline_options=pipeline_options
        )
    }
)

result = converter.convert("meeting_recording.wav")

​
AsrPipelineOptions

Configuration options for the Automatic Speech Recognition pipeline.

​
Parameters

​
asr_options
InlineAsrOptions
default:"asr_model_specs.WHISPER_TINY"
Automatic Speech Recognition (ASR) model configuration for audio transcription.Specifies which ASR model to use (e.g., Whisper variants) and model-specific parameters for speech-to-text conversion.
Inherits all parameters from PipelineOptions:
  • document_timeout
  • accelerator_options
  • enable_remote_services
  • allow_external_plugins
  • artifacts_path

​
Available Whisper Models

Docling provides pre-configured Whisper model options: 
from docling.datamodel import asr_model_specs

# Available models (smallest to largest)
WHISPER_TINY = asr_model_specs.WHISPER_TINY      # Fastest, ~1GB VRAM
WHISPER_BASE = asr_model_specs.WHISPER_BASE      # Balanced
WHISPER_SMALL = asr_model_specs.WHISPER_SMALL    # Good accuracy
WHISPER_MEDIUM = asr_model_specs.WHISPER_MEDIUM  # High accuracy
WHISPER_LARGE = asr_model_specs.WHISPER_LARGE    # Best accuracy, ~10GB VRAM

​
Model Selection Guide

Best for:
  • Quick drafts or previews
  • Resource-constrained environments
  • Real-time transcription
  • Simple audio with clear speech
Characteristics:
  • Size: ~75M parameters
  • Speed: Very fast
  • Accuracy: Basic
  • VRAM: ~1GB
Best for:
  • General-purpose transcription
  • Meetings and interviews
  • Balanced speed and accuracy
Characteristics:
  • Size: ~140M parameters
  • Speed: Fast
  • Accuracy: Good
  • VRAM: ~1-2GB
Best for:
  • Production transcription
  • Podcasts and lectures
  • Better accuracy needed
Characteristics:
  • Size: ~240M parameters
  • Speed: Medium
  • Accuracy: Very good
  • VRAM: ~2GB
Best for:
  • Professional transcription
  • Complex audio environments
  • Multiple speakers
Characteristics:
  • Size: ~760M parameters
  • Speed: Slow
  • Accuracy: Excellent
  • VRAM: ~5GB
Best for:
  • Maximum accuracy requirements
  • Difficult audio conditions
  • Multi-lingual content
  • Publication-quality transcripts
Characteristics:
  • Size: ~1.5B parameters
  • Speed: Very slow
  • Accuracy: Best
  • VRAM: ~10GB

​
Language Support

Whisper supports 90+ languages with automatic detection: 
# Automatic language detection (default)
pipeline_options = AsrPipelineOptions(
    asr_options=asr_model_specs.WHISPER_BASE
)

# Or specify language explicitly
# Language codes: en, es, fr, de, it, pt, nl, pl, ru, zh, ja, ko, etc.

​
GPU Acceleration

Enable GPU for faster transcription: 
from docling.datamodel.pipeline_options import AsrPipelineOptions
from docling.datamodel.accelerator_options import AcceleratorOptions

pipeline_options = AsrPipelineOptions(
    asr_options=asr_model_specs.WHISPER_BASE,
    accelerator_options=AcceleratorOptions(
        device="cuda",  # or "mps" for Apple Silicon
        num_threads=4
    )
)
See AcceleratorOptions for details.

​
Output Structure

Transcription appears as text items in the document: 
result = converter.convert("audio.mp3")

for item, _ in result.document.iterate_items():
    if isinstance(item, TextItem):
        print(f"Transcription: {item.text}")
        
        # Provenance may include timing information
        if item.prov:
            for prov in item.prov:
                print(f"  Timestamp: {prov.charspan}")

​
Advanced Usage

​
Batch Audio Processing

import concurrent.futures
from pathlib import Path

def transcribe_audio(audio_file):
    converter = DocumentConverter(
        format_options={
            AudioFormatOption: AudioFormatOption(
                pipeline_options=AsrPipelineOptions(
                    asr_options=asr_model_specs.WHISPER_BASE
                )
            )
        }
    )
    return converter.convert(audio_file)

# Process multiple audio files
audio_files = list(Path("recordings/").glob("*.mp3"))

with concurrent.futures.ThreadPoolExecutor(max_workers=2) as executor:
    results = list(executor.map(transcribe_audio, audio_files))

​
Extract from Video

# Video files automatically extract audio track
converter = DocumentConverter()
result = converter.convert("meeting.mp4")

# Transcription from video audio
print(result.document.export_to_text())

​
Long Audio Files

# For very long audio, use timeout protection
pipeline_options = AsrPipelineOptions(
    document_timeout=600.0,  # 10 minutes timeout
    asr_options=asr_model_specs.WHISPER_SMALL
)

​
Performance Optimization

Choose model based on requirements:
Use CaseModelSpeedAccuracy
Draft/PreviewTINY⚡⚡⚡⚡⭐⭐
GeneralBASE⚡⚡⚡⭐⭐⭐
ProductionSMALL⚡⚡⭐⭐⭐⭐
ProfessionalMEDIUM⚡⭐⭐⭐⭐⭐
Maximum QualityLARGE🐌⭐⭐⭐⭐⭐⭐
Enable GPU for significant speedup:
accelerator_options = AcceleratorOptions(
    device="cuda",  # 5-10x faster than CPU
)
Requirements:
  • NVIDIA GPU with CUDA support
  • Or Apple Silicon with MPS support
Larger models require more VRAM:
  • TINY/BASE: 1-2GB VRAM
  • SMALL: 2-3GB VRAM
  • MEDIUM: 5-6GB VRAM
  • LARGE: 10GB+ VRAM
For limited memory:
  • Use smaller model
  • Process on CPU (slower)
  • Chunk long audio files

​
Limitations

Known Limitations:
  • Processing time: Real-time factor varies by model (1x to 10x+ audio duration)
  • Accuracy: Depends on audio quality, accents, background noise
  • Speaker diarization: Limited speaker identification
  • Punctuation: May lack proper punctuation in some cases
  • Timestamps: Segment-level only, not word-level
  • Music/Sound effects: Not transcribed (speech only)

​
Troubleshooting

Solutions:
  1. Use larger model (MEDIUM or LARGE)
  2. Ensure audio is clear (reduce background noise)
  3. Check correct language is detected
  4. Verify audio format is supported
# Use larger model for better accuracy
pipeline = AsrPipelineOptions(
    asr_options=asr_model_specs.WHISPER_LARGE
)
Solutions:
  • Use smaller model (TINY or BASE)
  • Process on CPU instead of GPU
  • Split long audio into chunks
# Use CPU for large files
pipeline = AsrPipelineOptions(
    asr_options=asr_model_specs.WHISPER_TINY,
    accelerator_options=AcceleratorOptions(device="cpu")
)
Optimizations:
  • Enable GPU acceleration
  • Use smaller model
  • Process shorter segments
  • Use WHISPER_TINY for drafts
Solution: Convert audio to supported format (WAV, MP3, FLAC)
# Using ffmpeg
ffmpeg -i input.aac -ar 16000 output.wav

​
Best Practices

  1. Audio quality: Use high-quality recordings (16kHz+ sample rate)
  2. Model selection: Start with SMALL, adjust based on results
  3. GPU usage: Enable GPU for faster processing
  4. Batch processing: Process multiple files in parallel
  5. Timeout protection: Set document_timeout for long files
  6. Format conversion: Convert to WAV for best compatibility

​
Use Cases

Meeting Transcription

Convert recorded meetings to searchable text

Interview Processing

Transcribe interviews for qualitative research

Podcast/Lecture

Create text versions of audio content

Voice Notes

Convert voice memos to text

​
See Also

Build docs developers (and LLMs) love

Get started for freeTalk to us