AI Audio Providers — Research & Reference (TTS + Music Generation)¶
⚠️ Superseded in part (2026-07-31). The council-approved AI Generation Strategy now governs cloud provider selection, unit costs, caps (WING-299), and top-up packs (WING-300) — including the decisions that Suno BYOK is dead (no official API), ambient beds are CC0 field recordings rather than generated music, and cloud media lanes are Google paid APIs under indemnity-first. This doc is retained as the underlying BYOK/licensing research reference (archive-not-delete policy).
Purpose. WorkWingman is adding provider-agnostic audio features: text-to-speech narration / "audio overviews" of study material, and optional AI-generated music for study soundtracks and podcast intro/outro stings. This is a research/reference doc (not a two-track technical/plain doc). It exists so a later docs/technical/ design + a .NET adapter can be written with the licensing and pricing landmines already mapped.
Model. Tiered BYOK (bring-your-own-key), mirroring ApiKeyProviderRegistry (src/WorkWingman.Infrastructure/Integrations/ApiKeyProviderRegistry.cs). Every provider row is Id, DisplayName, WhatItUnlocks, SignupUrl, Kind, IsFree, FreeTierNote, ConfigKey, IsWired. Hard rules that already govern the registry apply here verbatim:
- A free/local default always works with no key. Key absent → that premium provider shows unavailable, free carries on.
- The app never buys keys, never signs up on the user's behalf. It deep-links the user to
SignupUrl; the user pastes the key; it lands in the KeePass vault. - Sandbox never falls back to appsettings/env keys (see
ApiKeyProvider.cs) — same guarantee must hold for audio adapters.
As-of date: 2026-07-07. Pricing and licensing in this space move fast (see the caveat at the very bottom). Every non-obvious claim is cited inline.
TL;DR — recommended tiers¶
| Tier | TTS pick | Music pick | Why |
|---|---|---|---|
| Tier 0 — free/local default (keyless) | Kokoro-82M (Apache-2.0, bundled ONNX) | Meta MusicGen only if run as a personal, non-shared default; otherwise no keyless music — see landmine | Kokoro sounds like a real person, commercial-safe, ships in-app. MusicGen weights are CC-BY-NC → cannot license shared output. |
| Tier 1 — affordable BYO-key | OpenAI gpt-4o-mini-tts (~$0.015/min) or Amazon Polly Neural ($16/1M) |
Stability Stable Audio 2.5 or Beatoven.ai (perpetual royalty-free license) | Cheap, real REST APIs, clear commercial terms on paid tier. |
| Tier 2 — premium BYO-key | ElevenLabs (best-in-class, cloning/emotion) | ElevenLabs Music v2 (licensed-only training) or Google Lyria 3 (Vertex) | Highest quality; ElevenLabs Music v2 trained on cleared catalog = safest premium music. |
Best keyless default TTS: Kokoro (Apache-2.0). Realistically it sounds natural and human on normal-length narration; Piper is the fallback for very low-end CPUs but sounds more mechanical on long passages. Best keyless/local music: MusicGen technically runs locally, but its CC-BY-NC weights make its output unusable in anything the user shares/publishes — treat local music as "personal preview only," and gate any shareable music behind a Tier-1 BYO-key provider.
Licensing landmines (read before proposing anything)¶
- 🚑 Coqui XTTS v2 weights = non-commercial (CPML), and Coqui is defunct — no one to sell you a commercial license. Do not use for any user-shared audio. Code (MPL-2.0) is fine; the weights are the trap. (promptquorum)
- 🚑 Meta MusicGen / AudioCraft weights = CC-BY-NC-4.0. Code is MIT; you cannot sell or publish music made from the released weights. Fine as a personal-only preview; never as the engine behind content a user distributes. (HF discussion)
- ⚠️ Suno / Udio free tiers are non-commercial and the provider retains ownership. Even paid Suno now grants only a perpetual license to exploit, not copyright, and its ToS disclaims that any copyright vests at all. Neither exposes a first-party developer API — third-party "Suno API"/"Udio API" resellers are unofficial. (Suno commercial, Udio ToS)
- ⚠️ Free consumer tools ≠ licensed output. Google MusicFX (Labs) and Suno/Udio free tiers do not hand a clean commercial license; only the paid Lyria via Vertex does. (Lyria/MusicFX)
- ⚠️ Piper active fork is GPL-3.0 (was MIT). Usable commercially but copyleft — must disclose modifications if you ship a modified binary. Prefer Kokoro (Apache-2.0) as the bundled default to avoid the copyleft question entirely. (Piper license)
- ⚠️ OpenAI TTS requires an AI-disclosure to end users — you must tell listeners the voice is AI-generated. Trivial for us (an "AI narration" label), but it's a ToS obligation, not optional. (OpenAI TTS guide)
- ℹ️ Voice cloning everywhere requires documented consent (ElevenLabs, Resemble, WellSaid; and US state laws — CA §3344, NY §50-51, TN ELVIS Act). We are not cloning voices for narration, so this is out of scope — but if a "clone my own voice" feature is ever added, consent capture is mandatory. (ElevenLabs consent)
A) Premium Text-to-Speech APIs (narration / audio overviews)¶
All prices are BYO-key, pay-as-you-go, per 1M characters unless noted. "Commercial output OK?" = can the end user publish/share audio they generate.
| Provider | Public REST API + auth | Quality / notable | Price (BYO-key) | Free tier | Commercial output OK? | Gotchas |
|---|---|---|---|---|---|---|
| ElevenLabs | Yes, xi-api-key header (docs) |
Best-in-class realism; instant voice cloning, emotion (v3), streaming, ~300–500ms round-trip, 30+ langs | $0.10/1M (Multilingual v2/v3); $0.05/1M (Flash/Turbo) (API pricing) | Yes but non-commercial; commercial needs paid Starter $5/mo+ (BIGVU) | Yes on paid plan | Cloning needs consent; ElevenLabs keeps a license to your uploaded voice data to train models (policy) |
| OpenAI TTS | Yes, Authorization: Bearer, POST /v1/audio/speech (docs) |
13 voices on gpt-4o-mini-tts (steerable by instructions); streaming; mp3/opus/aac/flac/wav/pcm |
gpt-4o-mini-tts ~$0.015/min; tts-1 $15/1M, tts-1-hd $30/1M (pricing) |
No standing free tier (usage-billed) | Yes | Must disclose voice is AI to end users (ToS) (guide) |
| Google Cloud TTS | Yes, POST /v1/text:synthesize, OAuth/service-account or API key (docs) |
Chirp 3 HD (newest), Neural2, Studio; SSML; mp3/LINEAR16/OGG_OPUS; 40+ langs | Neural2 $16/1M, Chirp3-HD $30/1M, Studio $160/1M (pricing) | 1M chars/mo free (Chirp3/Neural2/Studio); 4M for WaveNet/Standard | Yes | Auth is service-account JSON, heavier than an API key; BYO-key UX = "paste GCP key" |
| Azure Neural TTS | Yes, Speech SDK/REST, subscription key (pricing) | Large voice/lang catalog; Neural HD; SSML; custom neural voice (enterprise) | Neural $16/1M, Neural HD $22/1M, commitment tiers down to $7.50/1M (TextToLab) | 500K chars/mo free | Yes (prebuilt); Personal Voice is consent-gated / not for commercial | Custom Neural Voice adds endpoint-hosting + training fees |
| Amazon Polly (Neural) | Yes, AWS SDK/SigV4 (pricing) | Neural + Generative + Long-Form engines; SSML; mp3/ogg/pcm; solid, not top-tier realism | Neural $16/1M; Generative $30/1M; Long-Form $100/1M; Standard $4/1M | 1M Neural chars/mo for 12 mo; free tier has no commercial restriction | Yes | SigV4 auth (BYO = access key + secret, two fields) |
| Cartesia (Sonic) | Yes, API key (docs) | #1 naturalness claims, ~40–90ms latency, 40+ langs, streaming, laughter/emotion | ~$35/1M (Sonic 3) (eesel) | Free $0 (20K credits/mo, non-commercial); commercial = Pro $5/mo (pricing) | Yes on Pro+ | Real-time-agent-oriented; overkill for batch narration but great if we want live playback |
| Deepgram (Aura-2) | Yes, API key (pricing) | Sub-90ms latency, agent-focused; per-1K-char billing | $30/1M PAYG, $27/1M growth | $200 free credit | Yes | Fewer voices/langs than the big cloud vendors |
| PlayHT | Yes, API key | Large voice library, cloning | Creator $39/mo (50K words), Pro $99/mo (g2) | 5K words/mo (non-commercial) | Paid tiers only | Commercial rights restricted to paid (freetts matrix) |
| Rime | Yes, API key (pricing) | Conversational/"spoken-style" voices, low latency | $75/1M | 10K chars/mo free | Yes on paid | Niche voice style; not general narration first choice |
| WellSaid Labs | Yes (Studio + API) | Enterprise narration, studio-trained pro-talent voices | Creative $50/mo/user (awesomeagents) | Trial only | Yes but enterprise contract; some voices restrict political content | Built for enterprise; heavy for a BYO consumer |
| Resemble AI | Yes, API key | Deep emotional/voice-param control, cloning | Free 12.5K chars/mo; paid from ~$374/yr for 3M (typecast) | 12.5K chars/mo | Paid tiers | Cloning consent required |
| Hume (Octave) | Yes, API key (docs) | Emotionally expressive, prompted with natural-language emotion cues (no SSML) | $30/1M PAYG (pricing) | Free/Starter = non-commercial; commercial = Creator $14/mo | Creator+ only | Great for "warm study narrator" tone; commercial gate at $14/mo |
| Murf | Yes, API key (API plans) | Studio-quality + Falcon conversational | Studio $30/1M, Falcon $10/1M; $10/mo free API credit | $10/mo API credit | Yes (paid, Creator+) | Two API models at different quality/price |
Free / local tier (keyless, bundled)¶
| Engine | License | Quality reality | Commercial output? | Fit |
|---|---|---|---|---|
| Kokoro-82M | Apache-2.0 (HF) | "Sounds like a real person" — natural intonation, smooth prosody; 54 quality-graded voices, 8 langs; 82M params, runs on CPU (RTF ~0.03 on GPU); 24kHz WAV native (review) | Yes — fully commercial | ✅ Recommended Tier-0 default |
| Piper | GPL-3.0 (active fork; was MIT) | Clear but more mechanical on long passages; 900+ voices, 35+ langs; ~30–50ms first-audio, king of Raspberry-Pi/edge (codesota) | Yes (copyleft — disclose mods) | ✅ Fallback for very low-end CPUs |
| Coqui / XTTS v2 | Code MPL-2.0; weights CPML = NON-COMMERCIAL | Best open multilingual voice-cloning quality | 🚑 NO — non-commercial weights, vendor defunct | ❌ Do not use for shared audio |
| Windows SAPI | OS built-in | Robotic legacy voices | Yes (OS) | ⚠️ Last-resort only if no bundled model; poor quality |
B) AI Music Generation APIs (study soundtrack / podcast stings)¶
Emphasis: is there a real API (not just a web app)? and commercial/ownership terms.
| Provider | Real API? | Quality / notable | Price | Commercial / ownership | Verdict |
|---|---|---|---|---|---|
| ElevenLabs Music v2 | Yes — POST /v1/music (docs) |
Studio-grade; trained on licensed-only catalog; section inpainting, embedded SFX; mp3/pcm/opus | $0.15/min (API) (pricing) | Cleared for commercial on paid plan; licensed-only training = strongest legal footing | ✅ Best premium Tier-2 music |
| Stability Stable Audio 2.5 | Yes — Developer Platform API, credits (platform) | Trained on licensed AudioSparx catalog; instrumental focus | Credit-based (1 credit=$0.01), 25 free credits (stableaudio) | Commercial on paid; free access for orgs under $1M/yr revenue (license) | ✅ Good affordable Tier-1 |
| Google Lyria 3 (Vertex) | Yes — Vertex AI API (public preview) (docs) | Lyria 3 (30s clips) / Lyria 3 Pro (up to 184s, vocals) | Lyria 2 ~$0.06/30s (blog) | Paid Vertex = commercial; free MusicFX Labs does not grant a clean license | ✅ Premium alt (needs GCP) |
| Beatoven.ai | Yes — API | Mood/scene-based custom tracks for video | Subscription | Perpetual royalty-free license, survives cancellation (beatoven) | ✅ Creator-friendly Tier-1 |
| Mubert | Yes — API for apps/games/streams (api) | 200+ moods; loop/stream-oriented royalty-free | Subscription | Sync license for video/podcast/app/game (license) | ✅ Good for ambient study loops |
| Meta MusicGen / AudioCraft | Local (MIT code) + hosted on Replicate (replicate) | Text→music; medium model needs ~16GB GPU | Free (local) / Replicate per-run | 🚑 Weights CC-BY-NC — output NOT licensable for sharing (HF) | ⚠️ Local personal-preview only, never shared output |
| Stable Audio Open 1.0 | Local weights (HF) | Open self-host variant | Free (self-host) | Community license — check per-use; more permissive than MusicGen but read terms (HF) | ⚠️ Possible future local option; verify license before shipping |
| AIVA | Limited/enterprise | Composition tool | Free / Std / Pro | Copyright only transfers on Pro; Free/Std = AIVA keeps copyright (wavespeed) | ⚠️ Ownership gated to Pro |
| Loudly / Soundraw | Loudly has API; Soundraw web-first | Stem/loop hybrid, section reshaping; trained on own catalogs (sidesteps label disputes) | Subscription | Royalty-free on paid plans | ⚠️ OK Tier-1 alt; confirm API access |
| Suno | No first-party API (only unofficial resellers) | Top consumer quality | Consumer subscription | Free = non-commercial + Suno owns; paid = perpetual license, not copyright; ToS disclaims copyright vesting (dynamoi) | ❌ No official API — do not propose |
| Udio | No first-party consumer API; v4 "Enterprise API" exists but gated (musicmake) | Top consumer quality, 48kHz | Consumer subscription | Free = Udio owns; Pro/Enterprise = you own | ❌ Not a BYO-key consumer API — skip |
| Google MusicFX | No API (Labs web tool) | Free consumer demo | Free | No clean commercial license | ❌ Web-only, unlicensed — do not propose |
| Riffusion | Developer/customization niche API | Spectrogram-based | Varies | Commercial on paid | ℹ️ Niche; not a first pick |
Consumer-web-only (no usable BYO API) → do NOT propose: Suno (first-party), Udio (consumer), Google MusicFX.
C) Recommendation — tiered registry mapping¶
Concrete ApiKeyProviderRegistry rows to add (Kind = CredentialKind.ApiKey unless noted). IsWired = false until the adapter exists.
Tier 0 — free/local default (keyless, no registry row needed)¶
| Feature | Engine | Why |
|---|---|---|
| TTS narration / audio overview | Kokoro-82M bundled ONNX | Apache-2.0, natural voice, CPU-capable, 24kHz WAV — always-on default with zero key. |
| Music (personal preview only) | MusicGen local (optional, gated "not for shared content") | Runs locally but CC-BY-NC → never feed into user-published audio. If we don't want the licensing footgun, ship no keyless music and require a Tier-1 key for any shareable soundtrack. |
Tier 1 — affordable BYO-key¶
| Id | DisplayName | WhatItUnlocks | SignupUrl | IsFree | ConfigKey | Why |
|---|---|---|---|---|---|---|
tts-openai |
OpenAI TTS | Higher-quality narration (gpt-4o-mini-tts) |
https://platform.openai.com/api-keys | false | WorkWingman:OpenAiTtsApiKey |
~$0.015/min, one API key, clear commercial rights (must show AI-voice label). |
tts-polly |
Amazon Polly (Neural) | Cheap high-volume narration | https://aws.amazon.com/polly/ | false | WorkWingman:PollyAccessKey (+:PollySecretKey) |
$16/1M, 12-mo free tier with no commercial restriction. |
music-stableaudio |
Stability Stable Audio | Royalty-free study/intro music | https://platform.stability.ai/ | true | WorkWingman:StableAudioApiKey |
Licensed-catalog training; free for <$1M-rev users; cheap credits. |
music-beatoven |
Beatoven.ai | Mood-based royalty-free soundtracks | https://www.beatoven.ai/ | true | WorkWingman:BeatovenApiKey |
Perpetual license survives cancellation — safest creator terms. |
Tier 2 — premium BYO-key¶
| Id | DisplayName | WhatItUnlocks | SignupUrl | IsFree | ConfigKey | Why |
|---|---|---|---|---|---|---|
tts-elevenlabs |
ElevenLabs | Best-in-class narration, emotion | https://elevenlabs.io/app/settings/api-keys | true | WorkWingman:ElevenLabsApiKey |
Top realism; free tier exists but commercial needs Starter $5/mo. |
tts-google |
Google Cloud TTS | Chirp 3 HD narration, 40+ langs | https://console.cloud.google.com/apis/credentials | true | WorkWingman:GoogleTtsApiKey |
1M free chars/mo; reuses existing Google-Cloud BYO pattern. |
music-elevenlabs |
ElevenLabs Music | Cleared, licensed-only AI music | https://elevenlabs.io/app/settings/api-keys | true | WorkWingman:ElevenLabsApiKey (shared) |
$0.15/min; licensed-only training = strongest premium-music legal footing. |
music-lyria |
Google Lyria (Vertex) | Structured tracks w/ vocals | https://console.cloud.google.com/ | false | WorkWingman:GoogleTtsApiKey (shared GCP) |
Paid Vertex grants commercial license; Pro up to 184s. |
IsFree flag note: set IsFree = true where a genuinely free tier exists for the key itself (ElevenLabs, Google, Stable Audio, Beatoven all have free tiers, even if commercial use needs the cheapest paid plan — mirror the FMR/Udemy convention where IsFree means "free to obtain," with the caveat in FreeTierNote). Set IsFree = false for OpenAI/Polly/Lyria which are usage-billed with no standing free tier.
Bad-fit flags for user-shared content: Coqui/XTTS (non-commercial weights), MusicGen (CC-BY-NC weights), Suno/Udio free tiers (provider-owned, no official API), MusicFX (no license, no API). None of these should back any audio a user publishes.
D) Integration notes (for the later .NET adapter)¶
TTS #1 — Kokoro (Tier-0 local default)¶
- No HTTP. Load
kokoro-v1.0.onnx+voices-v1.0.binviaMicrosoft.ML.OnnxRuntime; output is a float sample buffer at 24000 Hz, write to WAV (NAudio/libsndfile). (kokoro-onnx, ONNX weights) - Pick voice by id (e.g.
af_sarah),speedfloat,lang="en-us". Batch (no streaming needed for overviews). Convert WAV→MP3 with a bundled encoder if smaller files wanted.
TTS #2 — OpenAI (Tier-1 batch)¶
POST https://api.openai.com/v1/audio/speech, headerAuthorization: Bearer <key>.- Body:
{ "model": "gpt-4o-mini-tts", "input": "<text>", "voice": "alloy", "response_format": "mp3", "speed": 1.0 }. response_format∈mp3|opus|aac|flac|wav|pcm. Response body = raw audio bytes (stream to file). Supports chunked streaming for play-while-generating. (docs)
TTS #3 — Google Cloud TTS (Tier-2 batch)¶
POST https://texttospeech.googleapis.com/v1/text:synthesize, auth via API key or service-account bearer.- Body:
{ "input": {"text": "..."}, "voice": {"languageCode":"en-US","name":"en-US-Chirp3-HD-..."}, "audioConfig": {"audioEncoding":"MP3"} }. audioEncoding∈LINEAR16|MP3|OGG_OPUS. Response JSON has base64audioContent(decode → file). Batch only (no low-latency stream for narration). (docs)
Music #1 — ElevenLabs Music (Tier-2)¶
POST https://api.elevenlabs.io/v1/music, headerxi-api-key: <key>.- Simple:
{ "prompt": "calm lo-fi study beat", "music_length_ms": 30000 }(music_length_ms3000–600000; can't combinepromptwithcomposition_plan). - Output format string
codec_sample_rate_bitrate, defaultmp3_48000_192; also PCM/opus/u-law/a-law. Streaming endpoint available (/v1/music/stream). (compose docs)
Music #2 — Stability Stable Audio (Tier-1)¶
- Stability Developer Platform REST endpoint, API-key auth, credit-based (1 credit=$0.01). Text-prompt → audio; instrumental focus. Confirm current endpoint path + max duration at build time. (platform)
Adapter shape. One IAudioNarrator seam (Kokoro default + OpenAI/Google/ElevenLabs BYO adapters) and one IMusicGenerator seam (Stable Audio/Beatoven/ElevenLabs/Lyria BYO adapters), each resolving its key through IApiKeyProvider.Get(id) exactly like the existing GeminiImageSource/OpenAiImageSource do. Absent key → adapter reports unavailable; Kokoro carries narration; music simply hidden. Sandbox never reads appsettings keys.
Subscriptions vs. API keys — what users may already have¶
The BYO-key screen asks the user to paste an API key. A common (and expensive) misconception is that a consumer subscription — ChatGPT Plus, Gemini Advanced, an ElevenLabs Creator plan — comes with a usable API key. For most providers it does not: the consumer chat/app subscription and the developer API are separate accounts with separate billing. This section web-verifies the 2026 reality so our UI copy never misleads a subscriber into thinking they're already set up. As-of 2026-07-07.
1. OpenAI — Plus/Pro subscription ≠ API key (confirmed, unchanged)¶
- A ChatGPT Plus ($20/mo) or Pro ($200/mo) subscription does NOT include API access, an API key, or API credits. It only upgrades the chat web/app interface. (OpenAI Help — billing in ChatGPT vs Platform)
platform.openai.comis a fully separate, pay-as-you-go account with its own billing. Even Pro/Business/Enterprise chat subscriptions give no API credit and no API discount. (ToolColumn — subscription vs API billing)- No standing free API tier. API usage is billed per token from the first call; you must add a payment method / prepay credits on the platform account. So a user with only a ChatGPT subscription has no usable key and must set up separate API billing to use our OpenAI TTS adapter.
2. Google — the free AI Studio key is the real win; the consumer sub is not it¶
Two very different things share the "Gemini" brand — keep them apart:
- (a) FREE Google AI Studio Gemini API key — go to aistudio.google.com, create a key, and get a genuinely free API tier: Gemini Flash / Flash-Lite at ~1,500 requests/day, 1M tokens/min, no credit card required. (Gemini API free tier, TokenMix — 1,500 req/day, no card) Caveat: in April 2026 Google moved Pro models to paid-only; the free tier is now Flash-only, and free-tier prompts may be used to improve Google's products. This is the key we already use for
gemini-image— the same AI-Studio key. - (b) Vertex AI (Chirp 3 HD TTS, Lyria 3 music) — these are NOT on the free AI Studio key. They require a Google Cloud project with billing enabled (service-account or GCP API key), billed per character / per 30s. (Lyria on Vertex)
- Consumer subscription grants NEITHER. Google AI Pro / Gemini Advanced / Google One AI Premium (~$19.99/mo) does NOT include API access — it's a consumer chat subscription (Gemini app, 100 prompts/day, Deep Research, etc.). It gives no Gemini API key and no Vertex access. (AIonX — Gemini Advanced API access, pecollective — free tier guide)
Bottom line for Google: point users at the free AI Studio key for anything that model supports; only the (separately-billed, GCP-project) Vertex path unlocks Chirp3 TTS / Lyria music — and never imply their $19.99 Gemini sub covers either.
3. ElevenLabs / Amazon / Stability / Beatoven¶
- ElevenLabs — subscription DOES include API (the one exception). From Starter ($5–6/mo) upward, API access is included on the same account, drawing from the same shared monthly credit pool as the web app — there's no separate API account. (ElevenLabs pricing, Flexprice breakdown) So an existing ElevenLabs paid subscriber already has a usable API key in their account settings. (Free plan can generate an API key too, but its output is non-commercial.)
- Amazon Polly — needs a separate AWS account (not Prime). There's no "Polly subscription." The user signs up for AWS (separate from their amazon.com/Prime account), creates an IAM access-key + secret, and is billed pay-as-you-go; a card is required even for the 12-month free tier. (AWS — signing up, Polly IAM) BYO here is two fields (access key + secret), not one.
- Stability AI — Developer Platform API is separate from the consumer "Stable Assistant" subscription. The API is pay-as-you-go credits ($10/1,000 credits after 25 free). A Stable Assistant chat/image subscription does not hand you a Platform API key. (Stability Developer Platform) (25 free credits let a user try our music adapter with no spend.)
- Beatoven — API is a separate, contact-gated key. Consumer plans (Creator ~$2.50/mo etc.) are for the web app; API access requires emailing
hello@beatoven.aito be issued an API key for your app. A normal subscriber does not automatically have an API key. (Beatoven API)
Honest messaging table (what the UI may safely tell the user)¶
For each provider: does something the user might already own give them a usable API key? And the exact one-liner we can show without lying.
| Provider | If you already have… | Usable key? | Safe one-liner for the UI |
|---|---|---|---|
| OpenAI TTS | ChatGPT Plus / Pro | NO | "Your ChatGPT subscription does not include an API key. Create a separate, pay-as-you-go key at platform.openai.com." |
| Google (Gemini API models) | nothing | YES — free | "Get a free Google AI Studio API key — no credit card, generous free tier." ⭐ |
| Google Chirp 3 / Lyria (Vertex) | Google AI Pro / One AI Premium | NO | "Your Gemini subscription doesn't cover this. Chirp 3 TTS / Lyria music need a Google Cloud project with billing enabled." |
| ElevenLabs | An ElevenLabs paid plan (Starter+) | YES | "Already subscribed to ElevenLabs? Your paid plan includes API access — copy your key from ElevenLabs → Settings → API Keys." |
| Amazon Polly | Amazon Prime | NO | "Prime doesn't count — Polly needs a separate AWS account. Create an IAM access key (pay-as-you-go; free tier available)." |
| Stability Stable Audio | Stable Assistant subscription | NO (but 25 free credits) | "The consumer app plan isn't the API. Grab a Developer Platform key — 25 free credits to start." |
| Beatoven | A Beatoven web-app plan | NO | "Beatoven's API key is issued on request — email hello@beatoven.ai to enable it for the app." |
Emphasis for the UI:
- ⭐ Genuinely free to grab (no card): the Google AI Studio Gemini API key — this is the one to surface first and label "Free." It's the only recommended provider where a brand-new user gets a working key at zero cost and no billing setup.
- 💳 Pay-as-you-go, separate billing (set expectations): OpenAI, Amazon Polly, Google Vertex (Chirp3/Lyria), Stability (after 25 free credits) — all require the user to create a developer account and add a payment method; their existing consumer subscription won't help.
- ✅ Subscription already includes the key: only ElevenLabs — an existing paid subscriber can reuse their key directly; everyone else must set up a distinct API/developer account.
- 🚫 Never claim a chat subscription = an API key. Specifically: a ChatGPT Plus/Pro user and a Google AI Pro / Gemini Advanced user have no usable API key from that subscription — the UI must not imply otherwise.
As-of 2026-07-07. Provider billing structures change; OpenAI has periodically floated (but not shipped) subscription-linked API perks, and Google's free-tier limits tightened in April 2026. Re-verify each provider's "subscription includes API?" answer before shipping the copy above.
Caveat — as-of 2026-07-07, what's likely to change¶
- Prices drop fast. ElevenLabs already cut API + Music pricing up to 50% in 2026; Azure Neural HD fell $30→$22; assume all per-1M figures are ceilings, re-check before quoting to users.
- Music ownership/licensing is legally unsettled. Suno/Udio terms shifted mid-2026 after label deals (Warner/Suno) — ownership language keeps changing and copyright vesting is explicitly disclaimed. Re-verify any music provider's commercial terms at integration time and again before any "publish your podcast" feature ships.
- Model names churn. Chirp 3 HD, Sonic 3, Lyria 3, Aura-2, ElevenLabs Music v2,
gpt-4o-mini-ttsare current today; expect version bumps. Pin the model id in config, not code. - Local model licenses can flip. Piper already went MIT→GPL-3.0; XTTS weights are stranded by Coqui's shutdown. Prefer Apache-2.0 Kokoro for the bundled default specifically because its license is the least likely to bite.
- Verify official endpoints at build time — several exact paths/params above come from vendor docs that revision frequently; treat this doc as the map, not the final contract.