JayrajLakkad/Emotion_Agnostic_Audio

Got it — here’s the same card with hyphen bullets and no trailing dots

---

license: cc-by-4.0 task_categories:

• automatic-speech-recognition
• audio-to-audio language:
• en tags:
• audio
• speech
• ASR
• emotion-suppression
• emotion-normalization
• monotonic-speech
• voice-conversion
• cyclegan
• world-vocoder
• ravdess pretty_name: 'Emotion-Agnostic Audio (EAA)' size_categories:
• 1K<n<10K

---

Dataset Card for Emotion-Agnostic Audio (EAA)

• Emotion-neutralized speech from RAVDESS
• Prosody flattened with WORLD
• Residual emotion reduced with EmoCycleGAN
• Linguistic content and speaker identity preserved
• Target use — robust ASR under expressive speech

Dataset Details

Dataset Sources [optional]

• Repository — add your Hugging Face dataset URL
• Paper [optional] — add preprint or DOI link
• Demo [optional] — add audio samples page

Uses

Direct Use

• ASR training and evaluation with emotion-agnostic inputs
• Preprocessing baseline where emotion is a confounder for speech analytics

Out-of-Scope Use

• Emotion recognition
• Prosody-sensitive TTS or style transfer that needs expressive speech

Dataset Structure

• Item fields

  • path — WAV mono 16 kHz recommended
  • speaker_id — 1–24
  • gender — {male, female}
  • original_emotion — {neutral, calm, happy, sad, angry, fear, disgust, surprise}
  • intensity — {normal, strong} if available
  • sentence_id — {1, 2}
  • processing_stage — {monotonic_world, neutralized_gan}
  • split — {train, validation, test} speaker-disjoint recommended

• Splits — 80/10/10 by speaker to prevent leakage

Dataset Creation

Curation Rationale

• Standardize prosody to isolate lexical and phonetic content for ASR robustness

Source Data

Data Collection and Processing

• Source — RAVDESS speech subset acted English
• Processing — WORLD-based F0 flattening then EmoCycleGAN neutralization plus optional loudness normalization and trimming
• Output — neutralized WAVs with metadata csv or json

Who are the source data producers

• Professional actors balanced male and female scripted sentences multiple emotions

Annotations [optional]

Annotation process

• Labels inherited from source filenames and metadata emotion intensity actor sentence
• Additional field processing_stage added automatically

Who are the annotators

• Not applicable no manual relabeling beyond metadata parsing

Personal and Sensitive Information

• No direct PII actors identified by numeric IDs avoid re-identification attempts

Bias, Risks, and Limitations

• Acted studio speech with limited lexical diversity typically two sentences
• Monotone processing reduces naturalness not suited for prosody studies
• English only accents and conditions limited to source corpus

Recommendations

• Report results on raw versus neutralized to show impact
• Use speaker-disjoint splits and publish preprocessing scripts and configs

Citation [optional]

BibTeX

@dataset{eaa_2025,
  title   = {Emotion-Agnostic Audio (EAA)},
  author  = {Lakkad, Parth and Collaborators},
  year    = {2025},
  url     = {https://huggingface.co/datasets/<your-namespace>/<dataset-name>}
}

APA

• Lakkad, P., & Collaborators 2025 Emotion-Agnostic Audio EAA Dataset Hugging Face

Glossary [optional]

• WER — word error rate equals substitutions plus deletions plus insertions divided by reference word count
• CER — character error rate defined analogously at character level
• WORLD — vocoder used for F0 estimation and manipulation
• CycleGAN — unpaired mapping used for emotion to neutral spectral conversion

More Information [optional]

• Provide preprocessing code configs metrics scripts and sample notebooks

Dataset Card Authors [optional]

• Parth Lakkad primary maintainer

Dataset Card Contact

• Email — your-email@example.com
• Issues — use the Community tab on the dataset page