Qwen3-TTS voices
Hindi and Odia, four named voices, one model. Pick a voice, then a training step. Clips are the 14 test sentences, not seen in training. Model and full details.
The reference clips come from earlier runs with other voices. Compare pronunciation and timing, not the voice itself. Untrained Qwen3 uses its stock voice and was never taught Odia. There are no error rates or listening scores for this model yet, so judge by ear. Test sentences are invented examples, not statements of fact, and all speech here is synthetic. Reference clips: IndicF5 by AI4Bharat (MIT) and Qwen3-TTS by the Qwen team (Apache 2.0).