fix(tts): improve Opus encoding quality for Telegram voice messages

[shellybotmoyer]

Summary

Fixes #18818 — suboptimal Opus encoding for Gemini/Edge TTS causing quality loss in Telegram voice messages.

Problem

Both _convert_to_opus() and _generate_gemini_tts() use ffmpeg with suboptimal Opus encoding settings:

ffmpeg -i input -acodec libopus -ac 1 -b:a 64k -vbr off output.ogg -y

Issues:

• -vbr off forces Constant Bitrate (CBR) — wastes bits on silence, starves complex segments
• Missing -application voip — Opus encoder defaults to audio mode, not optimized for speech
• Missing -compression_level 10 — lower compression quality

This causes:

1. Noticeably lower speech quality compared to properly-encoded Opus
2. Inconsistent Telegram voice bubble rendering (some files appear as regular audio instead of voice messages)

Fix

Changed ffmpeg encoding parameters at both call sites:

Parameter	Before	After	Why
-b:a	64k	48k	48k VBR is standard for VoIP; VBR uses bits more efficiently
-vbr	off (CBR)	on (VBR)	VBR allocates bits where needed for speech
-application	(missing)	voip	Optimizes Opus for voice intelligibility
-compression_level	(missing)	10	Maximum encoding quality

Testing

• ffmpeg -application voip is a standard libopus option (value 2048) that enables CELT+SILK hybrid mode optimized for speech
• 48k VBR is the standard bitrate for high-quality VoIP Opus
• No functional changes to the conversion pipeline — only encoding parameters changed

Affected Code

• tools/tts_tool.py:712 — _convert_to_opus() (used by Edge TTS, xAI TTS, and other providers)
• tools/tts_tool.py:1201 — _generate_gemini_tts() inline conversion

[teknium1]

Thanks for addressing both Telegram Opus conversion paths. The current-main premise is confirmed: tools/tts_tool.py:916 and tools/tts_tool.py:1795 still use -vbr off.

Problems

• The change has no regression coverage for the resulting ffmpeg argv. tests/tools/test_tts_opus_routing.py:47-70 mocks _convert_to_opus rather than its subprocess call, and tests/tools/test_tts_gemini.py currently has no ffmpeg/.ogg assertion.
• The same Telegram encoder settings remain duplicated in the two paths (tools/tts_tool.py:915-916 and tools/tts_tool.py:1792-1797). The linked fix(tts): optimize Opus output for Telegram voice #30066 discussion identifies a shared command builder plus tests as the more thorough form of this fix.

Suggested changes

• Add mocked-subprocess tests for both paths, asserting -vbr on, -application voip, and -compression_level 10.
• Centralize the shared Telegram Opus argv while retaining each caller’s existing subprocess options.

Automated hermes-sweeper review.

[teknium1]

Merged into main via consolidated salvage PR #73508 (merge f440a44753). Your Opus VBR/voip encode settings were cherry-picked with your authorship and applied to the current central transcoders.

Your contribution is credited to you in git history. Thank you! Closing this PR as merged-via-salvage.