Voice-Note-Audio / parameters.md
danielrosehill's picture
Initial commit
6158f1c
|
raw
history blame
4.08 kB

Annotation Parameters

This document outlines the key parameters and aspects to be annotated within the voice notes dataset. Annotations are added progressively as part of the review process.

Audio Quality

  • Clarity: How clear is the primary speaker's voice?
    • Options: [Very Clear, Clear, Somewhat Muffled, Muffled, Very Muffled]
  • Background Noise Level: The overall level of background noise.
    • Options: [None, Low, Moderate, High, Very High]
  • Noise Type: The type of background noise present.
    • Options: [None, Music, Crowd, Traffic, Nature, White Noise, Other (Specify)]
  • Reverberation: The level of echo or reverberation.
    • Options: [None, Low, Moderate, High]

Speaker Characteristics

  • Primary Speaker Dominance: How much of the audio is dominated by the primary speaker?
    • Options: [>90%, 70-90%, 50-70%, 30-50%, <30%]
  • Number of Speakers: An estimate of how many distinct speakers are present.
    • Options: [1, 2, 3, 4, 5+]
  • Speaker Overlap: How much do speakers talk over each other?
    • Options: [None, Minimal, Moderate, Frequent]

Transcription Quality

  • AI Transcript Accuracy (Overall): A general assessment of the Voicenotes.com STT accuracy for the entire note.
    • Options: [Very High (>95%), High (90-95%), Moderate (80-90%), Low (70-80%), Very Low (<70%)]
  • Specific Error Types: Check all that apply.
    • Incorrect Words
    • Missed Words/Phrases
    • Added Words/Phrases
    • Punctuation Errors
    • Capitalization Errors
    • Speaker Labeling Errors (if multiple speakers)
  • Difficult Segments: Note timecodes of segments that are particularly difficult for STT.

Context & Content

  • Primary Topic: The main subject of the voice note.
    • (Free text or predefined list based on your notes)
  • Language: The primary language of the audio.
    • (E.g., English, Spanish, etc.)
  • Technical Jargon: Is there specialized terminology?
    • Options: [None, Low, Moderate, High]

Recording Details

  • Recording Location: General location where the note was recorded.
    • (Text field, default: Jerusalem)
  • Recording Environment: Whether the recording was made indoors or outdoors.
    • Options: [Indoor, Outdoor]
  • Microphone Source: The general source of the microphone used.
    • Options: [Phone Internal Mic, Bluetooth Mic, Desktop Mic]
  • Microphone Type: The specific type of microphone used.
    • Options: [Earpiece, Lavalier, Gooseneck, Desktop]

Content Type

  • Voice Note Content Type: The primary purpose or type of content captured in the note. (Defined in label_studio_config.xml)
    • Options:
      • Blog Outline
      • Email Draft
      • Calendar Appointment
      • Note To Self
      • Reminder
      • Task List
      • Grocery List
      • Shopping List (Other)
      • Online Shopping
      • Stack Research
      • AI Prompt
      • System Prompt

Annotation State

  • Corrected: Whether the AI transcript has been manually corrected.
    • Options: [Yes, No, Partially]
  • Annotator Notes: Any additional observations or comments.

Auto-Computed Fields

The following metrics will be automatically calculated and added to the dataset metadata:

  • Run Time (of file): Duration of the audio file.
  • Word Count (of transcript): Total number of words in the transcript.
  • Character Count (of transcript): Total number of characters in the transcript.
  • Word Error Rate (of original AI transcript): Calculated WER comparing the AI transcript to the corrected transcript.
  • Estimated Speaker WPM: Estimated words per minute for the primary speaker.
  • Speaker WPM Classification: Classification of speaker speed (1-5).
  • Average dB Level: Average decibel level of the audio.

Related Datasets

This dataset is part of a broader STT fine-tuning project that will produce multiple related datasets:

  • Voice Note Audio (this dataset): Public dataset on Hugging Face for real-world STT evaluation
  • Basic STT Evaluation for Synthetic Voice Notes: Complementary dataset for controlled evaluation scenarios