ElevenLabs and Cartesia covered quality-default and latency-default TTS. PlayAI is the third generative-voice RFP name: the conversational/platform evolution of PlayHT (Play.ht). Vol. 16.2 still teaches PlayHT APIs; Vol. 22.4 teaches the company you sign after rebrand and agent packaging. After PlayAI the module leaves spoken voice and enters music: Suno.
Capability: TTS, cloning. Product: Vol. 21 voice assistants. Wrap: Vol. 18 SDKs / FastAPI.
Learning Objectives
By the end of this lesson, students should be able to:
- Map PlayAI to the PlayHT / Play.ht lineage without treating the rebrand as a new modality.
- Describe PlayAI as generative TTS, cloning, and conversational-voice platform (STT not the flagship).
- Compare PlayAI vs ElevenLabs vs Cartesia vs hyperscalers with qualitative pricing/strengths/weaknesses.
- Call TTS using a PlayHT-style or current PlayAI client and isolate it behind FastAPI.
- Design STT+TTS split: Whisper/Deepgram/hyperscaler in, PlayAI out.
- Apply the same consent and PII rules as any cloning vendor.
PlayAI is the voice-AI platform identity of Play.ht / PlayHT: cloud neural TTS, instant cloning, streaming synthesis (Play 3 / PlayDialog-class engines—names evolve), and conversational agents. Authentication historically used an API key plus user id (pyht-style). Any transcription features are secondary. For production ASR, pair PlayAI with a dedicated STT stack from Vol. 16.2 lectures 1–6.
Rebrand vs Capability
Procurement docs may say PlayHT, Play.ht, or PlayAI. Engineers should treat them as one vendor family unless a contract explicitly splits products. Do not rewrite Vol. 16.2 labs solely because the homepage title changed. Do unbundle: WER lives on the listen side; MOS and TTFB live on the speak side.
Pricing / Strengths / Weaknesses (Qualitative)
| Dimension | PlayAI / PlayHT | ElevenLabs | Cartesia | Hyperscaler TTS |
|---|---|---|---|---|
| Pricing posture | Typically character- or plan-based TTS; cloning and agent tiers—confirm live sheet | Characters/credits + STT minutes if used | Usage/plan Sonic TTS | Metered characters; enterprise commits |
| Strengths | Generative TTS + dialog engines; streaming; PlayHT API familiarity; cloning | Studio/brand/cloning reputation; agent UX | Live TTFB specialist | DPA, regions, STT+TTS one cloud |
| Weaknesses | STT not why you buy; name churn confuses RFPs; cloud; smaller “default” status than ElevenLabs in some markets | Price/TTFB trade-offs | Narrower studio narrative | Less “character” voice |
Engines and Product Shapes
Batch TTS
- Course narration, IVR prompts
- Job → download audio
- Cache aggressively
Streaming / PlayDialog-class
- Conversational turns
- gRPC/WS via SDK
- Barge-in still your gateway
Agents
- Vendor-hosted voice agent
- Fine for prototypes
- Production: your LLM + wrap-as-data
Pick PlayAI when
- Existing PlayHT integration / keys
- Dialog-class TTS wins listening tests
- You already own STT
- You want a generative alternative to ElevenLabs
Avoid using it as
- Your WER source of truth
- On-prem speech appliance
- A music or video generator
- A license to clone without consent
Python: PlayHT-Lineage TTS Call
PlayAI SDK package names may differ from classic pyht. Pattern below matches Vol. 16.2: user id + API key from env, engine/voice from config. Isolate behind FastAPI so switching to ElevenLabs is a client swap.
Related Lectures
| Lecture | Role |
|---|---|
| PlayHT | Capability + historic API |
| ElevenLabs (STT/TTS) / 22.4 ElevenLabs / Cartesia | Voice peers |
| Whisper / Deepgram | STT split |
| FastAPI / streaming | Product wrap |
| Suno | Next: music, not voice |
“PlayAI is a new modality.” It is PlayHT’s platform evolution. Second: PlayDialog means you can skip STT specialists. Third: user id in the PlayHT client is optional security theater—it is auth; treat it like a secret. Fourth: rebrand resets consent for old clones. Fifth: PlayAI competes with Suno on songs.
Knowledge Check
- Short Answer: What older product name should students map to PlayAI? Answer: PlayHT / Play.ht.
- True/False: PlayAI is primarily an STT specialist like Deepgram. Answer: False — TTS/cloning/agents first.
- Multiple Choice: Production ASR should usually be: (a) a dedicated STT vendor or Whisper, (b) implied by PlayDialog, (c) Suno lyrics. Answer: (a).
- Short Answer: Name one qualitative strength vs hyperscaler TTS. Answer: Generative/dialog voice quality, cloning, or streaming TTS (any valid).
- True/False: PlayHT user id + API key belong in source control. Answer: False.
- Multiple Choice: Vol. 16.2 PlayHT lecture is: (a) still the capability/API companion, (b) obsolete and must be ignored, (c) a video model. Answer: (a).
- Short Answer: Which module starts music generation after PlayAI? Answer: Suno (then Udio) in 22.4.
- True/False: Cloning consent resets automatically after a vendor rebrand. Answer: False.
- Multiple Choice: Isolate PlayAI behind: (a) FastAPI/your gateway, (b) a public HTML voice id, (c) Veo. Answer: (a).
- Short Answer: Name one TTS RFP peer in this module. Answer: ElevenLabs or Cartesia.
Key Takeaways
- PlayAI = PlayHT lineage generative TTS/agents; rebrand ≠ new science.
- Unbundle STT vs TTS; eval MOS/TTFB independently.
- Keep keys and voice IDs in env/config; wrap with Vol. 18.
- Same Vol. 20 consent/PII rules as ElevenLabs.
- Next: Suno—text-to-song, not TTS.
Lab: Given a PlayHT-style snippet and an ElevenLabs snippet, students write an adapter interface synthesize(text, voice_id) -> bytes with two implementations. Grade: no secrets in git, STT called out as a third interface, one paragraph on when PlayAI wins a listening test vs Cartesia TTFB.
Recap: PlayAI is the third generative-voice vendor, continuous with Vol. 16.2 PlayHT. Spoken-voice catalog done. Music next: Suno.