speechspectrum 1.0.0

Tools for obtaining speech representations along the speechspectrum.

🗣️ SpeechSpectrum

SpeechSpectrum provides a lightweight Python package for generating transcript representations along a linguistic fidelity spectrum—from verbatim transcripts to enhanced and compressed forms such as bullet points. The repository contains all code, notebooks, and the static site used in the accompanying user study.

⚙️ Requirements

• Python 3.8+
• OpenAI API credentials (API key + organization/project IDs)

📦 Installation

• Via PyPI:

  pip install speechspectrum
  

• From source (this repository):

  git clone https://github.com/mariateleki/SpeechSpectrum.git
  pip install -e .
  

🚀 Example Usage (also see notebooks/SpeechSpectrum_Tool.ipynb)

from speechspectrum import Client

client = Client(
    openai_api_key="sk-...",
    openai_org_id="org_...",      # pass "" if your account does not use orgs
    openai_project_id="proj_..."  # pass "" if your account does not require it
)

# Optional: begin from an audio file
# verbatim = client.audio_to_verbatim("path/to/audio.wav").text

MEDICAL_EXAMPLE = {
    "verbatim": (
        "Okay, so, um, the patient came in today and, uh, she was "
        "complaining of, you know, chest pain that's been going on for about, "
        "let's see, um, three days now. And, uh, she said it gets worse when she, "
        "when she breathes deeply or, or coughs. So I, I examined her and found "
        "some, uh, some tenderness in the, the inter – intercostal muscles on the "
        "right side. Her vital signs were, um, let me think, blood pressure was "
        "one-twenty over eighty, heart rate was, was seventy-two, and temperature "
        "was normal at, uh, ninety-eight point six. I'm thinking this is probably, "
        "you know, costochon – costochondritis rather than anything, anything more "
        "serious like a cardiac event."
    ),
    "non-verbatim": "",
    "enhanced": "",
    "bullet-points": ""
}

# Execute the pipeline on the medical example (or replace MEDICAL_EXAMPLE['verbatim']
# with audio_verbatim above to process an observed recording)
MEDICAL_EXAMPLE["non-verbatim"] = client.verbatim_to_nonverbatim(MEDICAL_EXAMPLE["verbatim"])
MEDICAL_EXAMPLE["enhanced"] = client.nonverbatim_to_enhanced(MEDICAL_EXAMPLE["non-verbatim"])
MEDICAL_EXAMPLE["bullet-points"] = client.enhanced_to_bulletpoints(MEDICAL_EXAMPLE["enhanced"])

print("Verbatim:", MEDICAL_EXAMPLE["verbatim"])
print("Non-verbatim:", MEDICAL_EXAMPLE["non-verbatim"])
print("Enhanced:", MEDICAL_EXAMPLE["enhanced"])
print("Bullet points:", MEDICAL_EXAMPLE["bullet-points"])

📂 Repository Structure

• src/speechspectrum/ — Source code for our Python package, using gpt-4o-transcribe (audio → verbatim) and gpt-5.1 (downstream text transformations).
• notebooks/SpeechSpectrum_Tool.ipynb — Shows end-to-end transcript pipeline with results used in the website + user study.
• docs/ — Static website served at https://speechspectrum.org.
• notebooks/run_llm_experiment.ipynb — Executes the persona-based prompting study modeling preference distributions across fidelity levels.
• notebooks/get_figs_and_sig_tests.ipynb — Reproduces statistical analyses and figure generation from the study outputs.

🌐 Demo Website

The static demonstration resides in docs/. The site is published via GitHub Pages at https://speechspectrum.org.

🧭 Usage Principles and Ethical Considerations

SpeechSpectrum is designed to support user agency and transparency in transcript generation. Key principles for responsible use include:

• Context-dependent fidelity: Different tasks require different levels of detail. Select fidelity settings based on the informational and situational needs of the domain.

• Transformations are interpretive: Cleaning, enhancing, or summarizing transcripts alters meaning. These operations can suppress uncertainty cues or standardize informal or dialectal speech.

• Linguistic representation matters: Post-processing can unintentionally erase identity markers or stylistic features. Exercise care when working with marginalized or diverse speech communities.

• Disfluencies are informative: Filled pauses, hesitations, and repairs can carry semantic or pragmatic value. Preserve or remove them thoughtfully, especially in high-stakes contexts.

• Protect speaker privacy: Audio and transcripts may encode sensitive information. Collect and process data with consent and in accordance with relevant privacy requirements.

• Use LLM preference modeling cautiously: LLM personas tend to overgeneralize or exaggerate role-based preferences. Empirical user data should guide design decisions.

• Maintain transparency: When sharing outputs, disclose the fidelity level used and any post-processing applied so downstream users can interpret transcripts appropriately.

📚 Citation

@inproceedings{choi2025speechspectrum,
  title     = {SpeechSpectrum: A Framework for Speech-to-Text Representation Along the Linguistic Fidelity Spectrum},
  author    = {Choi, Anna Seo Gyeong and Teleki, Maria and Miguel del Rio and Corey Miller and James Caverlee and Allison Koenecke},
  booktitle = {Under Review},
  year      = {2025}
}

📜 License

MIT, see LICENSE.