Skip to content

Voice cloning that doesn’t sound cloned

A Hindi & English text-to-speech studio built on a 6 GB laptop GPU, and the engineering it took to make the cloned voice sound natural.
⚡ My first GPU project · NVIDIA RTX 3050 Laptop · 6 GB VRAM · 35 W

Same text, same voice, same GPU. The only difference is the pipeline. Press play, then flip between Before and After: the playhead stays in place, so you hear the same moment both ways.

English

“Every voice tells a story…”

BeforeOriginal pipeline
  • 19.1s
  • 930 ms dead silence
  • noise floor -64 dB
AfterImproved pipeline
  • 20.95s
  • 10 ms dead silence
  • noise floor -67.2 dB
0:00Switch Before/After while playing; the position is kept so you can compare the same moment.
2.9 s0.13 s
Dead digital silence
in a 69 s narration, measured on the client voice
0.9330.944
Speaker similarity
model’s own voice encoder, same seeds
≈1.3×
Real-time on 6 GB
1 s of audio in ≈1.3 s once warm
3.8 GB
Peak GPU memory
fits a 6 GB laptop GPU with room to spare

Text → natural speech

Paste up to 30,000 characters of Hindi, Marathi (Devanagari) or English. It is chunked, synthesized and stitched into a single narration.

Zero-shot voice cloning

A 10–30 s reference recording is enough. The pipeline cleans it and picks the best 10 s for the model.

Runs on a laptop

Built and measured on an RTX 3050 6 GB laptop GPU, with one resident model and a single-GPU job queue.

Responsible by default

Consent is required for custom voices. Every output carries a watermark and AI-generated metadata tags.

  1. 1
    Script cleanupCPU

    Punctuation, abbreviations, markdown, numbers

  2. 2
    Smart chunkingCPUimproved

    Sentence & Danda (।) aware, ≤220 chars; mid-sentence cuts keep a comma

  3. 3
    Voice conditioning clipCPUimproved

    De-clip, de-rumble, light denoise, pick the best 10 s window

  4. 4
    Chatterbox synthesisGPU

    Resident model, cached voice conditioning, tuned settings

  5. 5
    Chunk checkCPUimproved

    Babble, truncation and dead-air detection; one retry with a new seed

  6. 6
    Natural joiningCPUimproved

    Trim, level-match, pauses by punctuation, filled with room tone

  7. 7
    Polish & labelCPUimproved

    Two-pass loudness to −16 LUFS + AI-attribution tags in one encode

  8. 8
    DeliverCPU

    Exactly two files: WAV master + MP3