Shxbhxm21/lecture-tts-audio

Lecture TTS Audio (Hinglish)

Synthesized lecture narration for Indian school lessons. Each lecture's teacher speech is rendered to audio with the VoxCPM2 + Hinglish LoRA voice. Speech text is in pure Devanagari (English technical terms spelled phonetically in Devanagari, e.g. polygon → पॉलीगॉन), and every audio clip is paired back to its text and to every place it occurs in the lecture script.

• Audio: 48 kHz mono WAV
• Voice/model: VoxCPM2 + Hinglish LoRA (step 100), cfg 2.0, 10 diffusion timesteps

Lectures

mensuration-class8/

• Title: Mensuration — Area of a Polygon (splitting into triangles & trapeziums)
• Curriculum: CBSE Class 8 Math — NCERT Chapter 11: Mensuration, Section 11.1
• Clips: 98 unique audio files · 164 speech occurrences · ~27.9 min total

Layout

<lecture>/
  lecture_audio/<hash>.wav     one wav per UNIQUE speech string (md5(text)[:12])
  manifest.json                [{audioFile, text, durationSec, sampleRate, occurrences[]}]
  lecture_with_audio.json      full lecture script with `audioFile` next to each `speech`
  lecture_devanagari.json      the Devanagari lecture script (no audio paths)

Using the manifest

manifest.json is the index. One entry per unique clip:

{
  "audioFile": "lecture_audio/cb8c942cff83.wav",
  "text": "नमस्ते बच्चों। ... एरिया ऑफ़ अ पॉलीगॉन। ...",
  "durationSec": 11.04,
  "sampleRate": 48000,
  "occurrences": [
    {"jsonPath": "root.beats[0].segments[0].speech",
     "beatId": "intro-why-split-polygons", "segmentType": "greeting"}
  ]
}

Identical speech is deduplicated: one wav is referenced from every occurrence (e.g. the beats[] script and its slides[].beatMeta mirror share the same audio). Audio filenames are md5(text)[:12].wav, so re-generating the same text is idempotent.