Hindi TTS in Nkomo: Why Google Neural2 Still Leads Swara

For Hindi, Nkomo is deliberately not making the same change as Arabic and Mandarin. Google `hi-IN-Neural2-A` remains first. Microsoft `hi-IN-SwaraNeural` follows through Edge and then Azure.

The reason is simple: the evidence does not justify declaring a winner. Both are modern, credible Hindi voices. Online comments often call Google clear and consistent but restrained, while Swara is sometimes described as warmer or more naturally Indian. Those impressions vary with the passage, speaking rate and listener. They are not a native Nkomo benchmark.

What is actually being compared

Google's catalogue classifies Neural2 as a general-purpose neural tier. Microsoft lists Swara in its Hindi voice support. Nkomo uses the same Swara voice ID through Edge and Azure, so switching between those transports changes availability and operating cost, not the intended speaker.

Hindi also exposes a weakness hidden by polished demo sentences: real conversations are frequently Hinglish. English product names, addresses, abbreviations and amounts appear inside Hindi syntax. Devanagari, Romanized Hindi and mixed scripts can produce very different pronunciation from the same provider. A voice that wins on a literary Hindi paragraph may lose on the messages Nkomo actually receives.

The decision

The route remains Google Neural2, then Edge/Swara, then Azure/Swara. This is the conservative order because Google's configured voice is already a newer tier than the WaveNet models Nkomo had selected for Arabic and Mandarin, and the available public opinion is mixed.

The Edge transport comes before Azure after Google fails. The open-source edge-tts client reaches Microsoft's online Edge speech service without an API key. That makes it useful, but it is not a substitute for Azure's supported service contract. Keeping Azure behind it gives Nkomo a controlled fallback when the Edge endpoint changes, throttles or becomes unavailable.

What would change our mind

We need a blind test designed around Hindi as people use it: formal Hindi, everyday conversation, Hindi-English code-switching, Romanized input, names from different regions, dates, rupee amounts, phone numbers, questions and short assistant replies. Native reviewers should score intelligibility, naturalness, code-switch handling and fatigue over longer passages.

If Swara wins that test consistently, Nkomo can move Edge and Azure ahead of Google. Until then, keeping Neural2 first is not indecision. It is refusing to turn scattered preference into false certainty.