Add SLNG as Voice AI Infrastructure Provider

[LukeAM]

Add SLNG as Voice AI Infrastructure Provider

Summary

This PR adds SLNG as a provider to models.dev, contributing 21 TTS/STT model entries across regional voice AI infrastructure.

What is SLNG?

SLNG is a regional voice AI infrastructure company that operates voice AI models (Fish Audio, Deepgram, Cartesia, Soniox, and other TTS/STT models) across regions where model labs don't have their own infrastructure — primarily EU and AU, expanding to India, Japan, and other regions.

Models are offered in two modes:

• Self-hosted: SLNG runs the model on our own GPU infrastructure
• Proxied: SLNG routes requests to the model lab's API with regional optimizations

What's Included

Provider

• providers/slng/provider.toml — Provider configuration with OpenAI-compatible endpoint
• providers/slng/logo.svg — SLNG brand logo

Models (21 total)

• 13 Text-to-Speech models supporting 130+ languages across 9+ regions
• 8 Speech-to-Text models supporting 53+ languages

Represented Model Labs

• Deepgram (Aura 2, Nova 3)
• Fish Audio (TTS S2/S2.1 Pro)
• Cartesia (Sonic 3, 3.5)
• Soniox (TTS RT v1, Speech AI RT v5)
• Inworld (Max 1.5)
• Sarvam AI (Bulbul v3, Saaras)
• KugelAudio (Kugel 2)
• Murf (Falcon)
• Gradium (TTS, STT)
• Speechmatics (Realtime v2)
• Reson8 (STT v1)

Source of Truth

All data sourced from SLNG's authoritative model catalog: https://docs.slng.ai/models

• Model names, capabilities, and language support match SLNG docs exactly
• Pricing converted from per-minute to per-token format using industry-standard conversion rates:
  • TTS: 195 tokens/min (150 words/min × 1.3 tokens/word)
  • STT: 2,000 audio tokens/min
• No invented or extrapolated data

Schema Compliance

• ✅ All required fields present (name, modalities, release_date, last_updated, cost, limits)
• ✅ TOML syntax validated against existing provider patterns
• ✅ Logo uses currentColor and responsive viewBox (no fixed dimensions)
• ✅ OpenAI-compatible provider setup with API base URL
• ✅ Modalities correctly specified for TTS (text → audio) and STT (audio → text)

Testing

Validated against:

• Google Gemini TTS models (modality and cost patterns)
• OpenAI Whisper models (STT patterns)
• OpenRouter/DeepInfra provider patterns (inference provider structure)
• 5+ existing provider TOMLs for consistency

Models Excluded

Two models from the SLNG catalog are excluded from this PR:

• Kugel 1 (TTS) - Agent-bundle pricing only, no API pricing
• Fish STT - Agent-bundle pricing only, no API pricing

These may be added in a future update if API pricing becomes available.

Notes

1. Token limits: Standard defaults used (32k/16k for TTS, 0/0 for STT) as SLNG docs don't specify token limits
2. Release dates: Placeholder dates pending specific release information from SLNG team
3. Regional variants: Single price point used per model (some have regional variation in source docs)

Checklist

• Provider metadata added (provider.toml)
• Logo added in SVG format with currentColor
• All API-priced models from source catalog included
• Pricing converted to per-million-token format
• Source URLs documented in TOML comments
• No compliance claims without source (GDPR, SOC2, etc.)
• No individual voice IDs listed
• Modalities match model capabilities

Downstream Impact

SLNG models will be available in:

• opencode
• RubyLLM
• ReqLLM
• Symfony AI
• Arcanum
• Other tools consuming models.dev API

---

Submitted by: Luke Miller on behalf of SLNG
Contact: [Provide contact email for questions]
Documentation: https://docs.slng.ai

[github-actions]

Action items

• [critical] [violation] providers/slng/models/aura-2.toml:5 - Check: Non-lab hosts must use base_model; add complete models/<lab>/… metadata when missing. Why: SLNG is a multi-lab voice gateway (Deepgram, Fish Audio, Cartesia, Soniox, Sarvam, Inworld, Murf, Speechmatics, etc.), not the model creator. All 21 entries are full standalone definitions with no base_model, which AGENTS.md treats as a merge blocker. Nameable lab models (e.g. Deepgram Aura 2 / Nova 3, Fish S2/S2.1 Pro, Cartesia Sonic 3/3.5, Sarvam Bulbul/Saaras) must live under models/<lab>/ first. Action: For each nameable lab model, add a complete models/<lab>/<id>.toml, point every SLNG file at it with base_model, and keep only provider-specific fields (cost, real ID/name deltas, deployment notes). Do not ship full inline third-party copies.
• [high] [possible mistake] providers/slng/models/aura-2.toml:8 - Check: release_date / last_updated must be real model dates, not placeholders. Why: Every model uses release_date = "2024-01-01" and last_updated = "2025-01-01". Peers already disagree (e.g. Vercel Fish Audio S2.1 Pro is 2026-07-28). Identical placeholder dates make catalog chronology wrong for all 21 models. Action: Replace with documented lab/provider release and update dates (or drop dates only after moving shared facts onto accurate lab models/ entries).
• [medium] [possible mistake] providers/slng/models/aura-2.toml:22 - Check: TTS/STT [limit] values must match real API limits or established audio peers. Why: All TTS files set context = 32_768 / output = 16_384 with no cited token windows. Existing TTS/STT peers (OpenAI TTS/Whisper, Fish Audio on Vercel, Mistral Voxtral TTS) use context = 0 / output = 0 when token windows do not apply. These look invented. Action: Verify SLNG’s real limits; if none, use 0/0 like peer audio models, or document genuine limits from SLNG docs.
• [medium] [possible mistake] providers/slng/models/aura-2.toml:18 - Check: Costs must be accurate USD/MTok (or omit when billing is not token-based). Why: Pricing is converted with fixed 195 tok/min (TTS) and 2000 tok/min (STT), STT is often capped at $3.00, Fish is labeled “$/M UTF-8 bytes” but stored as token input/output, and TTS sets identical input and output. That can misstate real $/MTok for cost tooling. Mistral omits TTS cost when character billing does not map cleanly. Action: Confirm SLNG’s billed unit and publish true USD/MTok (or input_audio/output_audio if that is how SLNG meters); if only per-minute/character/byte pricing exists, document the conversion with rate/date in a leading comment—or omit cost when the schema unit cannot represent it—and stop inventing caps/symmetric I/O without evidence.