Bilingual Hindi-English TTS Training
Budget: ₹1,500 – ₹12,500 INR
I already have 1800 clean text-audio pairs that mix Hindi and English exactly in the way shown below, and I now want a fully trained text-to-speech model I can run locally for further research and fine-tuning.
file001|Socho ek aisi agency jo कभी news में openly नज़र नहीं आती…|Socho ek aisi agency jo कभी news में openly नज़र नहीं आती…
file002|उन्नीस सौ बासठ के Sino-India war के बाद भारत ने external intelligence की अहमियत समझी।|उन्नीस सौ बासठ के Sino-India war के बाद भारत ने external intelligence की अहमियत समझी।
…and so on up to file1800.
Key points you should know before you dive in:
• The system must understand context in both languages so that pronunciation, code-switching and emphasis feel natural rather than stitched together.
• Output quality needs to be natural and human-like, not robotic. Think conversational YouTube narration.
• You are free to bring in additional open-source bilingual or monolingual datasets to augment training, provided the licences are compatible and you document what you used.
Deliverables
1. A reproducible training pipeline (Python notebooks or scripts) that takes my 1800 lines plus any extra public data you add, then outputs a ready-to-use model checkpoint.
2. Inference script that accepts a mixed-language sentence and produces a WAV or MP3 locally without a heavy cloud dependency.
3. README explaining environment setup, hyper-parameters, and how to continue fine-tuning with new data.
4. Short demo clip generated by the model to confirm naturalness and correct code-switch handling.
Acceptance criteria
• Pronunciation accuracy above 95 % on a held-out set I will provide.
If you have hands-on experience with Tacotron 2, VITS, FastPitch, Coqui TTS or similar frameworks, you will find everything you need ready to go. I will supply the 1800 WAV files and corresponding CSV once we start.
file001|Socho ek aisi agency jo कभी news में openly नज़र नहीं आती…|Socho ek aisi agency jo कभी news में openly नज़र नहीं आती…
file002|उन्नीस सौ बासठ के Sino-India war के बाद भारत ने external intelligence की अहमियत समझी।|उन्नीस सौ बासठ के Sino-India war के बाद भारत ने external intelligence की अहमियत समझी।
…and so on up to file1800.
Key points you should know before you dive in:
• The system must understand context in both languages so that pronunciation, code-switching and emphasis feel natural rather than stitched together.
• Output quality needs to be natural and human-like, not robotic. Think conversational YouTube narration.
• You are free to bring in additional open-source bilingual or monolingual datasets to augment training, provided the licences are compatible and you document what you used.
Deliverables
1. A reproducible training pipeline (Python notebooks or scripts) that takes my 1800 lines plus any extra public data you add, then outputs a ready-to-use model checkpoint.
2. Inference script that accepts a mixed-language sentence and produces a WAV or MP3 locally without a heavy cloud dependency.
3. README explaining environment setup, hyper-parameters, and how to continue fine-tuning with new data.
4. Short demo clip generated by the model to confirm naturalness and correct code-switch handling.
Acceptance criteria
• Pronunciation accuracy above 95 % on a held-out set I will provide.
If you have hands-on experience with Tacotron 2, VITS, FastPitch, Coqui TTS or similar frameworks, you will find everything you need ready to go. I will supply the 1800 WAV files and corresponding CSV once we start.