TTS Voice Over Skill

Generates per-slide WAV voice-over files from YAML speaker_notes using Azure Speech SDK with SSML pronunciation control, or a locally installed Piper engine.

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.

Two engines are available through --engine:

• azure (default) sends SSML to Azure AI Speech neural voices. Use it for published narration.
• piper is an explicit optional local choice using a separately installed Piper executable. It needs no credentials or network access after the voice is downloaded. Narration never leaves the host. HVE demo-material CI does not use either speech engine; it generates silent videos instead.

Prerequisites

• Azure Speech resource — Free tier provides 500K characters per month.
• Authentication — Key-based (SPEECH_KEY) or Microsoft Entra ID (SPEECH_RESOURCE_ID).
• Region — SPEECH_REGION is required for Azure synthesis and has no default. Dry-run mode and the piper engine do 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. Operators must set an approved region and avoid sending regulated or confidential narration.

Key-Based Auth

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

Microsoft Entra ID Auth

Requires a custom domain on the Speech resource and Cognitive Services Speech User role.

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

Install dependencies:

# run from this skill folder
uv sync

Piper Engine

Piper is not a dependency of this skill. It is licensed GPL-3.0-or-later, so the skill invokes it as an external executable, the same way other skills invoke FFmpeg or LibreOffice. Install it and download a voice separately:

uv tool install piper-tts
uvx --from piper-tts python -m piper.download_voices en_US-norman-medium --data-dir ~/.local/share/piper
export PIPER_DATA_DIR=~/.local/share/piper

To run Piper without installing it, also set:

export PIPER_COMMAND="uvx --from piper-tts piper"

Variable	Purpose
PIPER_COMMAND	Command that runs Piper, split without a shell (default: piper)
PIPER_DATA_DIR	Directory holding downloaded voices; --piper-data-dir takes precedence

Voice models carry their own licenses and dataset provenance in each voice's MODEL_CARD. The default en_US-norman-medium model card records that it was trained from scratch on public-domain LibriVox recordings, and the Piper voices repository is MIT licensed. Check the model card before publishing narration from any other voice, because some voices have noncommercial or research-only source restrictions.

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

Generate voice-over offline with Piper:

uv run scripts/generate_voiceover.py --engine piper --collapse-newlines \
  --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

Parameter	Type	Default	Description
--dry-run	flag	false	Print SSML (azure) or plain text (piper) without generating audio
--engine	string	azure	Synthesis engine: azure or piper
--voice	string	en-US-Andrew:DragonHDLatestNeural	Voice name; the piper default is en_US-norman-medium
--rate	string	+10%	Azure speech prosody rate; ignored by piper
--piper-data-dir	path	PIPER_DATA_DIR	Directory holding downloaded Piper voices
--content-dir	path	content	Path to slide content directory
--output-dir	path	voice-over	Path to WAV output directory
--slide	integer	all slides	Repeat to synthesize only the selected canonical slide numbers
--lexicon	path	(auto-detect)	Custom acronyms.yaml path
--collapse-newlines	flag	false	Collapse newlines and whitespace runs in speaker notes into single spaces before synthesis
--verbose / -v	flag	false	Enable verbose (DEBUG) logging output

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.

Parameter	Type	Default	Description
--input	path	(required)	Source PPTX file path
--audio-dir	path	voice-over	Directory with slide-NNN.wav
--output	path	*-narrated.pptx	Output PPTX file path
--verbose / -v	flag	false	Enable verbose (DEBUG) logging output

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.

The piper engine has no SSML support, so it substitutes aliases directly into the text. Aliases written as spaced single letters are hyphenated first (H V E Core becomes H-V-E Core), because Piper reads hyphenated letters as one fluent run and spaced letters as slow separate words.

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

Issue	Solution
Set SPEECH_KEY ... or SPEECH_RESOURCE_ID	Export SPEECH_KEY (key auth) or SPEECH_RESOURCE_ID (Entra ID) with SPEECH_REGION.
SPEECH_REGION must be set	Export SPEECH_REGION with an approved Azure region. Synthesis has no default region.
Multiple WAV files map to slide N	Remove the duplicate WAV files (for example slide-1.wav beside slide-001.wav) and rerun embed_audio.py.
401 with Entra ID auth	Verify custom domain on the Speech resource and Cognitive Services Speech User role. RBAC propagation takes up to 5 minutes.
Empty WAV files or skipped slides	Verify speaker_notes: is present and non-empty in content.yaml.
Mispronounced acronyms	Add entries to acronyms.yaml with phonetic aliases.
azure-cognitiveservices-speech package is required	Run uv sync in the skill directory.
Piper executable not found	Install Piper separately or set PIPER_COMMAND, for example uvx --from piper-tts piper.
Piper cannot find the voice model	Download the voice with piper.download_voices and pass its directory through --piper-data-dir or PIPER_DATA_DIR.
Audio icon visible in PPTX	Reposition or resize the audio object in PowerPoint after embedding.
Authored slide animations missing after embedding	embed_audio.py replaces existing p:timing with narration timing; re-apply animations in PowerPoint after embedding audio.
Slides no longer advance on click after embedding	embed_audio.py sets advClick="0" for auto-advance. To re-enable, select all slides in PowerPoint and check Advance Slide > On Mouse Click in the Transitions tab.
Video export shows "No timings recorded"	Re-embed audio with the updated embed_audio.py which adds narration timing XML automatically.