Voice AI Integration Engineer

Builds speech-to-text pipelines converting audio into timestamped, speaker-attributed transcripts and subtitles.

2|Updated May 21, 2026
One-click install
npx skills add https://github.com/tcvdog/agency-agents-hermes --skill voice-ai-integration-engineer-tcvdog
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: Voice AI Integration Engineer
Source: https://github.com/tcvdog/agency-agents-hermes/tree/main/engineering/voice-ai-integration-engineer
Command: npx skills add https://github.com/tcvdog/agency-agents-hermes --skill voice-ai-integration-engineer-tcvdog

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires faster-whisper, pyannote.audio, torch, httpx.

What problem does it solve? Turning raw audio recordings into clean, structured, usable text is error-prone: unvalidated formats silently degrade accuracy, long files overflow model context windows, and speaker labels get lost before downstream systems ever see them. This Skill designs and builds end-to-end transcription pipelines that handle ingestion, preprocessing, chunking, transcription, diarization, and structured delivery without silent failures. ## Core Features & Use Cases - Audio Preprocessing & Validation: Probe files with ffprobe, resample to 16kHz mono, normalize loudness with ffmpeg, and chunk long recordings with overlap to prevent word splits. - Transcription & Diarization: Run faster-whisper or cloud ASR services (AssemblyAI, Deepgram, AWS Transcribe), then merge pyannote speaker diarization for speaker-attributed segments. - Structured Output & Integration: Export SRT/VTT subtitles and stable-schema JSON, then hand off to CMS platforms (Drupal, WordPress), REST APIs, GitHub Actions, or LLM summarization agents. - Use Case: You have a 90-minute multi-speaker podcast recording. Use this Skill to preprocess the audio, transcribe it with speaker labels, generate broadcast-compliant SRT subtitles, and post a structured transcript to your CMS. ## Quick Start Transcribe the attached meeting recording into a speaker-attributed transcript with SRT subtitles and a structured JSON file ready for summarization.

Frequently Asked Questions about Voice AI Integration Engineer

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I transcribe long audio files with Whisper?▼

Split long recordings into overlapping chunks (e.g., 30 minutes with 30-second overlap) using ffmpeg, transcribe each chunk with faster-whisper, then trim overlap regions during assembly. Never rely on the model's maximum input duration, since overflow silently corrupts output.

faster-whisper vs cloud ASR services like Deepgram or AssemblyAI?▼

faster-whisper runs locally with CTranslate2 optimization, suiting privacy-sensitive or offline content at lower cost. Cloud services like Deepgram and AssemblyAI offer built-in diarization and higher accuracy for high-volume batch work, but audio leaves your environment.

How do I add speaker labels to a Whisper transcript?▼

Run speaker diarization with pyannote.audio to get timed speaker segments, then assign each transcript segment the speaker with maximum time overlap. Passing a known speaker count to pyannote significantly improves diarization accuracy.

Why is my Whisper transcription accuracy suddenly worse?▼

The most common cause is unvalidated input: stereo audio, wrong sample rates, or video containers passed directly to the model. Resample to 16kHz mono with ffmpeg (-ar 16000 -ac 1) and normalize loudness before transcription to restore accuracy.

Can Whisper handle HIPAA or GDPR-sensitive audio?▼

Cloud ASR services route audio outside your environment, so regulated content should use local Whisper models like faster-whisper or whisper.cpp. Add PII detection and redaction as a configurable pipeline stage and enforce retention policies.

How do I generate SRT subtitles from a transcript?▼

Enable word-level timestamps during transcription, then export segments with formatted HH:MM:SS,ms timestamps and optional speaker prefixes. Validate reading speed against the broadcast standard of 20 characters per second and split long segments.