# Voice comparison, S02 ACL (English), same 73-word script

Dr Aman's feedback: voice okay, but pauses odd and no expression. Everything below is MEASURED (no listening possible). Pauses = ffmpeg silencedetect (-38 dB, 70 ms) at punctuation; pitch = Praat F0; accuracy = faster-whisper small.en vs script.

| Variant | Engine | Voice | Narration s | Full-stop pause mean (ms) | Comma pause mean (ms) | Pauses at punctuation | Pitch SD (semitones, whole / per word) | Loudness variation (dB SD) | Word error % | Cost | Licence |
|---|---|---|---|---|---|---|---|---|---|---|---|
| R0-current | edge-tts (current, +5%) | en-IN-PrabhatNeural | 31.53 | 534 (min 393) | 232 (min 165) | stops 8/8, commas 5/5 | 4.66 / 3.66 | 1.96 | 1.4 | Rs 0 | unclear for commercial use |
| A1-prabhat | A: edge-tts tuned (0%, per-sentence, gaps, pitch) | en-IN-PrabhatNeural (male) | 33.64 | 535 (min 413) | 393 (min 330) | stops 8/8, commas 5/5 | 4.66 / 3.68 | 1.87 | 1.4 | Rs 0 | unclear for commercial use |
| A2-neerja-expressive | A: edge-tts tuned | en-IN-NeerjaExpressiveNeural (female) | 32.08 | 606 (min 522) | 463 (min 435) | stops 8/8, commas 5/5 | 3.74 / 2.46 | 3.86 | 4.1 | Rs 0 | unclear |
| A3-neerja | A: edge-tts tuned | en-IN-NeerjaNeural (female) | 33.06 | 632 (min 559) | 432 (min 364) | stops 8/8, commas 5/5 | 4.35 / 2.41 | 2.63 | 1.4 | Rs 0 | unclear |
| B1-gemini25-charon | B: Gemini 2.5 Flash TTS (style prompt) | Charon (male) | 29.96 | 310 (min 15) | 120 (min 0) | stops 7/8, commas 3/5 | 4.59 / 4.02 | 2.2 | 0.0 | ~$0.008 (~Rs 0.65) | licensed for commercial use under Google terms |
| B2-gemini25-orus | B: Gemini 2.5 Flash TTS (style prompt) | Orus (male) | 29.2 | 404 (min 292) | 140 (min 81) | stops 8/8, commas 4/5 | 3.85 / 3.28 | 1.76 | 0.0 | ~$0.008 (~Rs 0.65) | licensed for commercial use under Google terms |
| B3-gemini38-charon | B: Gemini 3.8 Flash TTS (style prompt) | Charon (male) | 29.94 | 691 (min 462) | 272 (min 146) | stops 8/8, commas 5/5 | 5.58 / 5.3 | 2.9 | 2.7 | ~$0.010 (~Rs 0.85) | licensed for commercial use under Google terms |

C. ElevenLabs: SKIPPED. Neither ELEVENLABS_API_KEY line works: the first returns 401 invalid_api_key, the second is an API key ID not a key (400). No credit spent. Owner-side fix: create a fresh key (also see the duplicate-key note in memory).

D. Avatar voice: the Flow avatar test clip voice was made on the Mac and is not available here; the avatar hook carries its own voice.

## Findings

- The CURRENT voice (R0) already has measurable pauses (about 530 ms at full stops, 230 ms at commas) and a normal pitch spread. So the complaint is probably uniform, same-length pauses between every line (including mid-thought) and flat delivery, not missing punctuation. Pitch SD is almost the same for every edge-tts variant: per-sentence pitch offsets barely change edge-tts prosody (A1 vs R0: 4.66 vs 4.66 semitones). Edge-tts cannot be made much more expressive.
- A1 (edge Prabhat tuned): 150 ms comma pauses were inserted by cutting audio at commas, plus 300-380 ms at stops; all 13 punctuation points paused. Voice and expression remain the old flat edge voice, narration 2 s longer. The female Neerja voices are not suited to a male surgeon; Neerja Expressive has the most loudness variation among edge voices but a higher error rate.
- Gemini 2.5 Orus (B2): every full stop paused (min 292 ms), 4 of 5 commas, pause lengths vary naturally (80-620 ms), 0 percent word errors, shortest (29.2 s). Charon (B1) has more pitch variation but skipped the pause at one full stop and at 'hour,' so it is less consistent. Gemini 3.8 Charon (B3) has the most pitch and loudness variation but pauses up to 1.1 s, slow, 2.7 percent errors (misread words), so not used.
- Not measurable: warmth, accent authenticity. He should listen.

## Recommendation

Pick B2 (Gemini 2.5 Flash TTS, Orus) if he likes how it sounds: natural pauses, varied pitch, commercially licensed, about Rs 0.65 per 30 s clip. Keep A1 only as the free fallback. ElevenLabs needs a working key before it can be compared.

Total paid spend this run: 3 Gemini requests, about $0.026 (roughly Rs 2.2). Nothing published. Renders: S02-v3-voiceA.mp4 (A1, 36.7 s), S02-v3-voiceB.mp4 (B2, 32.2 s); captions re-timed to each voice's word timings (A: edge-tts boundaries, B: faster-whisper word timestamps aligned to the script). Build folder: projects/aman-samples-v3-S02-voice/ (voices/ holds every variant's wav and timings).
