tts-voiceover

TTS Voice Over Skill

Generates per-slide WAV voice-over files from YAML speaker_notes using Azure AI Speech neural voices with SSML pronunciation control.

Overview

This skill reads content.yaml files from a PowerPoint skill content directory, extracts speaker_notes fields, applies acronym aliases for correct pronunciation of technical terms, and produces one WAV file per slide. Supports dry-run mode for SSML template verification without Azure credentials.

Synthesis uses Azure AI Speech neural voices, including the HD voices such as en-US-Andrew:DragonHDLatestNeural. HD voices are offered in a subset of Azure regions; check the Speech service regions before choosing SPEECH_REGION.

Narration produced by this skill is synthetic. Tell listeners that the voice is AI-generated, following Microsoft's disclosure design guidelines for synthetic voices. HVE demo-material CI does not synthesize speech; it generates silent videos.

Prerequisites

• Azure Speech resource: see Azure AI Speech pricing for the free-tier allowance and HD voice rates.
• Authentication: Microsoft Entra ID (SPEECH_RESOURCE_ID, recommended) or a resource key (SPEECH_KEY).
• Region: SPEECH_REGION is required for synthesis and has no default. Dry-run mode does not need it.
• Python 3.11+ with uv for virtual environment management.
• Data handling note: Speaker-notes content is transmitted to the configured SPEECH_REGION for synthesis. For prebuilt neural voices, Microsoft states that neither the input text nor the output audio is stored in Microsoft logs; see Data, privacy, and security for text to speech. Operators must still set an approved region and avoid sending regulated or confidential narration.

Microsoft Entra ID Auth (Recommended)

Entra ID avoids storing a long-lived key. It requires a custom domain on the Speech resource and the Cognitive Services Speech User role for the signed-in identity. The script resolves the identity through DefaultAzureCredential, so an az login session, a managed identity, or a GitHub Actions OIDC login through azure/login all work.

export SPEECH_RESOURCE_ID="/subscriptions/.../Microsoft.CognitiveServices/accounts/your-resource"
export SPEECH_REGION="eastus"

Key-Based Auth

export SPEECH_KEY="your-speech-key"
export SPEECH_REGION="eastus"

When both SPEECH_KEY and SPEECH_RESOURCE_ID are set, the script warns and uses the key. Unset SPEECH_KEY to use Entra ID.

Install dependencies:

# run from this skill folder
uv sync

Quick Start

Verify SSML templates without generating audio:

uv run scripts/generate_voiceover.py --dry-run --content-dir path/to/content

Generate voice-over WAV files:

uv run scripts/generate_voiceover.py --content-dir path/to/content --output-dir voice-over

Embed audio into a PPTX deck:

uv run scripts/embed_audio.py --input deck.pptx --audio-dir voice-over --output deck-narrated.pptx

Parameters Reference

For different speakers, synthesize one voice's selected slides per invocation using repeated --slide <number> arguments. Only those canonical slide-NNN folders are read and their WAVs replaced; other speakers' outputs are untouched. Omit --slide to synthesize the whole deck.

generate_voiceover.py

embed_audio.py

Embeds WAV files into corresponding PPTX slides and adds narration timing XML so PowerPoint recognizes the audio for video export via File > Export > Create a Video > Use Recorded Timings and Narrations.

Each WAV file maps to a slide by the number in its name, so slide-1.wav and slide-001.wav both map to slide 1. When two files map to the same slide, embedding stops with an error and writes no output. WAV files numbered past the last slide are ignored with a warning. An unreadable WAV file is reported, its slide is left unchanged, and the remaining slides are still embedded.

Script Reference

Generate with custom voice and rate:

uv run scripts/generate_voiceover.py \
  --content-dir content \
  --output-dir voice-over \
  --voice "en-US-Jenny:DragonHDLatestNeural" \
  --rate "+5%"

Use a custom lexicon:

uv run scripts/generate_voiceover.py \
  --content-dir content \
  --lexicon custom-acronyms.yaml

Collapse newlines in speaker notes (recommended for block-scalar | notes, whose line breaks are otherwise spoken as pauses):

uv run scripts/generate_voiceover.py \
  --content-dir content \
  --collapse-newlines

Embed generated audio:

uv run scripts/embed_audio.py \
  --input slide-deck/presentation.pptx \
  --audio-dir voice-over \
  --output slide-deck/presentation-narrated.pptx

Acronym Lexicon

The lexicon controls SSML <sub alias> replacements for acronyms and technical terms. Create an acronyms.yaml file:

acronyms:
  HVE-Core: "H V E Core"
  OWASP: "Oh wasp"
  SBOM: "S Bomb"
  SLSA: "Salsa"
  CI/CD: "C I C D"

Lexicon resolution order:

1. Path specified via --lexicon argument.
2. acronyms.yaml in the content directory.
3. Built-in defaults covering common technical acronyms.

SSML Template

Each slide produces an SSML document:

<speak version="1.0" xmlns="http://www.w3.org/2001/10/synthesis"
 xmlns:mstts="http://www.w3.org/2001/mstts" xml:lang="en-US">
  <voice name="en-US-Andrew:DragonHDLatestNeural">
    <prosody rate="+10%">
      Text with <sub alias="Oh wasp">OWASP</sub> aliases applied.
    </prosody>
  </voice>
</speak>

Integration with PowerPoint Skill

This skill reads from the PowerPoint skill's content directory structure:

content/
├── slide-001/
│   └── content.yaml    # Must include speaker_notes: field
├── slide-002/
│   └── content.yaml
└── ...

Each content.yaml should contain a speaker_notes: field with the narration text. The generated WAV files are named slide-NNN.wav matching the directory names.

Troubleshooting