Borrowing it
Nothing to install: this file belongs to drakulavich/kesha-voice-kit. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.
curl -O https://raw.githubusercontent.com/drakulavich/kesha-voice-kit/main/.claude/skills/tts-internals/SKILL.mdgit clone --depth 1 https://github.com/drakulavich/kesha-voice-kitWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/drakulavich/kesha-voice-kit/tts-internals)<a href="https://agentmods.dev/skills/drakulavich/kesha-voice-kit/tts-internals"><img src="https://agentmods.dev/badge/skills/drakulavich/kesha-voice-kit/tts-internals/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/drakulavich/kesha-voice-kit/tts-internals"><img src="https://agentmods.dev/badge/skills/drakulavich/kesha-voice-kit/tts-internals.svg" alt="Reviewed on agentmods" width="80" height="20"></a>- NVIDIA SkillSpector warn
SkillSpector: 1 finding, up to high
These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →
- high Anti-Refusal · line 81 Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.Fix: Remove instructions that suppress warnings, disclaimers, or ethical commentary. Let the agent surface safety-relevant caveats to the user.
What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00104 | $0.03122 |
| Opus 5 | $0.00052 | $0.01561 |
| Sonnet 5 | $0.00021 | $0.00624 |
| Haiku 4.5 | $0.00010 | $0.00312 |
Grade A, and why
tts-internals scanned grade A with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 7d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Makes network callslowCapability
Not a fault in itself. Listed so you know the mod talks to something, and to what.
- `KESHA_MODEL_MIRROR` — redirect HF downloads to an internal mirror (#121), preserving `/<owner>/<repo>/resolve/<ref>/<file>` for `wget --mirror`; empty/unset = no-op. Rust `models/download.rs::apply_mirror` and TS `sta How it starts
The opening of the file, as written. The whole thing — 104 lines — stays where its author put it; the contents beside it link to each section on GitHub.
TTS internals
Engines
Text-to-speech via three engines selected by voice id prefix:
en-*→ Kokoro-82M. Separate model + per-voice style embedding. Output 24 kHz.ru-*→ Vosk-TTS (alphacep/vosk-tts). Multi-speaker model, 5 baked-in speakers. Output 22.05 kHz.macos-*→ AVSpeechSynthesizer Swift sidecar (#141). Zero model download, notification-grade quality, darwin-arm64 release feature setcoreml,tts,system_tts;kesha installplacessay-avspeech-darwin-arm64next to the engine and runtime lookup is sibling-first (rust/src/tts/avspeech.rs::helper_path).
Install Kokoro + Vosk-TTS explicitly with kesha install --tts (~990 MB). macos-* voices use installed macOS voices and need no model install.
Behavior, G2P, and SSML
- TTS models are never auto-downloaded —
kesha sayfails loudly with akesha install --ttshint when models are missing. kesha saywrites WAV mono f32 to stdout unless--outis given. Stderr is progress/errors only.- G2P split: English (
en/en-us/en-gb) uses embeddedmisaki-rs(Kokoro-trained inventory, no system deps, OOV letter-spell); Russian uses Vosk-TTS internals (BERT prosody + dictionary, no system deps);es/fr/it/ptuse CharsiuG2P on ONNX builds and FluidAudio's own G2P on darwin-arm64system_kokoro(see below); every other language bails out ofg2p::text_to_ipa_cachedwith a pointer to #212 — darwin-arm64hi/ja/zhnever reach that function at all (#492, below). espeak-ng (#210) was retired in #214, which also dropped CharsiuG2P (#123) once it had no callers left; CharsiuG2P came back for the Romance languages in #509, closing #212. - Auto-routing: omitted
--voicecalls TSNLLanguageRecognizerand picksen-am_michael,macos-com.apple.voice.compact.ru-RU.Milenaon darwin Russian, orru-vosk-m02elsewhere. Confidence < 0.5 or unmapped language falls to engine default. Routing table:src/voice-routing.ts::pickVoiceForLang. - SSML (
--ssml):ssml-parser; supports required<speak>root and<break time="...">; rejects<!DOCTYPE>; unknown tags (<emphasis>,<prosody>,<phoneme>,<say-as>) warn once and strip tags while synthesizing contained text.tts::ssml::parsereturnsVec<Segment>;tts::say()loads the engine once, concatenates text/silence f32 samples, then callswav::encode_wav. Scope/future tags: #122.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 7d ago Changed 0cb7995d7057
- 11d ago First seen · 104 lines · 104 tokens per session scan A acfd71159622
tts-internals is a skill published in the GitHub repository drakulavich/kesha-voice-kit (73 stars, last pushed yesterday), licensed MIT. It adds 104 tokens to every session and 3,122 once invoked, about $0.0005 per session on Opus 5. A static security scan graded it A with 1 finding (makes network calls). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
benchmark
Run performance + quality benchmarks. ASR reports WER, RTF, process memory, and throughput across engines/variants. Arguments include asr, tts, vad, diarize, asr-quick.
review-pr
Review a pull request for conceptual fit, architecture impact, adversarial failure modes, security risk, docs impact, regression risk, test coverage, and merge readiness. Use when asked to review a PR, check whether a PR is safe to merge, decide if more tests are needed, perform adversarial or security review, or…
yw-transcribe
A workflow for transcribing one local audio or video file into a traceable Chinese or mixed Chinese-English transcript, with optional timestamps and subtitle files.
whisper
OpenAI's general-purpose speech recognition model. Supports 99 languages, transcription, translation to English, and language identification. Six model sizes from tiny (39M params) to large (1550M params). Use for speech-to-text, podcast transcription, or multilingual audio processing. Best for robust, multilingual…
whisper
OpenAI's general-purpose speech recognition model. Supports 99 languages, transcription, translation to English, and language identification. Six model sizes from tiny (39M params) to large (1550M params). Use for speech-to-text, podcast transcription, or multilingual audio processing. Best for robust, multilingual…
voice-settings
Choose Cadence Code's local speech and transcription models from the Codex or Antigravity UI. Use only when the user explicitly invokes $voice-settings, /voice-settings, or asks to open Cadence Code settings.