Migrate from PlayHT to Inworld Realtime TTS After Shutdown

What Happened to PlayHT?

1. July 2025: Meta acquired the PlayAI team (roughly 35 people), absorbing them into Meta Superintelligence Labs. Meta confirmed the acquisition publicly on July 12, 2025.
2. Late July 2025: the PlayHT API went offline ahead of schedule, around July 26, 2025. Cartesia noted publicly at the time that "their API is already offline, and the full platform sunsets on December 31st," and platforms like Kore.ai discontinued PlayHT TTS support.
3. December 31, 2025: all products (the play.ht studio, the API, Voice Agents) closed permanently. User accounts, saved audio, and voice clones were deleted at sunset with no export tooling.
4. As of July 7, 2026: the play.ht domain no longer resolves.

• Existing PlayHT API integrations are dead. The endpoints have been offline since late July 2025.
• Voice clones are gone. PlayHT deleted user voice data at the December 31, 2025 sunset. There is no recovery path through PlayHT.
• Your only option is to reclone from your original audio sources into a new provider.

Model Mapping: PlayHT to Realtime TTS

Migration Step 1: Reclone Your Voices

import requests
import base64

with open("original_voice_sample.wav", "rb") as f:
    audio_b64 = base64.b64encode(f.read()).decode()

# Step 1: Clone the voice
clone_response = requests.post(
    "https://api.inworld.ai/voices/v1/voices:clone",
    headers={"Authorization": "Basic <your-api-key>"},
    json={
        "displayName": "Customer Service Agent",
        "langCode": "EN_US",
        "voiceSamples": [
            {"audioData": audio_b64}
        ],
        "audioProcessingConfig": {"removeBackgroundNoise": True}
    }
)
voice_id = clone_response.json()["voice"]["voiceId"]

# Step 2: Use the cloned voice in TTS calls
tts_response = requests.post(
    "https://api.inworld.ai/tts/v1/voice",
    headers={"Authorization": "Basic <your-api-key>"},
    json={
        "text": "Hello, how can I help you today?",
        "voiceId": voice_id,
        "modelId": "inworld-tts-2",
        "audioConfig": {
            "audioEncoding": "MP3",
            "sampleRateHertz": 24000
        }
    }
)

with open("output.mp3", "wb") as f:
    f.write(base64.b64decode(tts_response.json()["audioContent"]))

• 5-15 seconds of clean, single-speaker original audio (samples >15s auto-trimmed).
• Formats: WAV, MP3, WEBM. Max 4MB per sample.
• Cloned-voice storage scales with plan: 100 custom voices on the free tier up to 30,000 on Growth, per inworld.ai/pricing as of July 7, 2026 (higher limits via enterprise sales).
• Use original human-recorded audio, not AI-generated PlayHT output. Generation-on-generation cloning compounds artifacts.

Migration Step 2: Swap the Synthesis API

# Before: PlayHT synthesis (legacy reference)
# response = playht.tts(text=text, voice="s3://voice-cloning-zero-shot/...", ...)

# After: Realtime TTS synthesis
import requests
import base64

response = requests.post(
    "https://api.inworld.ai/tts/v1/voice",
    headers={"Authorization": "Basic <your-api-key>"},
    json={
        "text": "Hello world",
        "voiceId": "Sarah",  # or your cloned voiceId
        "modelId": "inworld-tts-2",
        "audioConfig": {
            "audioEncoding": "MP3",
            "sampleRateHertz": 24000
        }
    }
)

audio_bytes = base64.b64decode(response.json()["audioContent"])
with open("output.mp3", "wb") as f:
    f.write(audio_bytes)

import requests
import base64
import json

with requests.post(
    "https://api.inworld.ai/tts/v1/voice:stream",
    headers={"Authorization": "Basic <your-api-key>"},
    json={
        "text": "Hello world",
        "voiceId": "Sarah",
        "modelId": "inworld-tts-1.5-mini",  # mini for lowest time-to-first-audio
        "audioConfig": {
            "audioEncoding": "PCM",
            "sampleRateHertz": 24000
        }
    },
    stream=True
) as r:
    for line in r.iter_lines():
        if not line:
            continue
        chunk_obj = json.loads(line)
        audio_bytes = base64.b64decode(
            chunk_obj["result"]["audioContent"]
        )
        # play / forward audio_bytes to client

Migration Step 3: Voice Library Mapping

import requests

response = requests.get(
    "https://api.inworld.ai/voices/v1/voices?languages=EN_US",
    headers={"Authorization": "Basic <your-api-key>"}
)

for voice in response.json()["voices"]:
    print(voice["voiceId"], voice["displayName"], voice.get("description"))

Migration Checklist

• [ ] Pull all original human-recorded voice samples used for PlayHT clones.
• [ ] Get an Inworld API key from platform.inworld.ai.
• [ ] Reclone each voice via POST /voices/v1/voices:clone. Save the returned voiceId.
• [ ] Update synthesis calls: change endpoint, use the voiceId and modelId field names, add audioConfig (audioEncoding + sampleRateHertz). If you are new to the API, the Python TTS API tutorial covers the full request/response cycle.
• [ ] Update streaming parsing: PlayHT used various streaming formats; Realtime TTS streaming is NDJSON with base64 audioContent per line.
• [ ] Decode base64 before writing audio. Both sync and streaming responses return base64.
• [ ] Test latency end-to-end. TTS 1.5 Mini delivers ~120ms median time-to-first-audio; TTS-2 and 1.5 Max deliver sub-200ms.
• [ ] If you used PlayHT for the full speech loop, consider migrating the LLM layer too. The Router is OpenAI SDK-compatible and routes to 220+ models at cost (OpenAI, Anthropic, Google, Groq, Fireworks, Mistral, DeepSeek and more).

Why Realtime TTS After PlayHT

• Voice quality. Inworld's Realtime TTS-2 is the #1 realtime TTS. PlayHT was a strong product; Realtime TTS-2 (research preview) pushes further on expressiveness, conversational naturalness, and natural-language steerability, with Realtime TTS 1.5 Max and Mini GA for production workloads. Judge it with your own ears: reclone one voice on the free tier and A/B it against your archived PlayHT output.
• 15 GA languages, plus 90+ experimental on TTS-2. TTS-2 preserves a cloned voice's identity across languages, so one reclone covers your whole language footprint instead of one clone per locale.
• Full pipeline integration. Realtime TTS pairs with Realtime STT, the Router (220+ models, at cost), and the Realtime API for end-to-end voice applications, so PlayHT teams that had stitched TTS to a separate LLM and STT stack can consolidate onto one provider and one API key.

When Another Provider Is the Better Fit

• You need a massive off-the-shelf voice marketplace or studio-first dubbing workflows. ElevenLabs has a large community voice library and mature dubbing/audiobook tooling. If that was why you chose PlayHT, evaluate ElevenLabs alongside Inworld; and if you later leave it, the ElevenLabs migration guide follows the same reclone-and-swap pattern as this one. Price it honestly: as of July 7, 2026, ElevenLabs lists Flash/Turbo at $0.05 per 1K characters ($50 per 1M) and Multilingual v3 at $0.10 per 1K ($100 per 1M) on elevenlabs.io/pricing/api, versus $25 per 1M on-demand for TTS-2 on inworld.ai/pricing.
• Your workload is long-form narration rendered offline, where time-to-first-audio is irrelevant and per-character price at batch volume dominates. Compare batch-oriented options in the best TTS APIs comparison before defaulting to a realtime-optimized engine.
• You are locked into a single-vendor end-to-end voice model stack (for example OpenAI's Realtime API) and do not want a cascaded pipeline. You trade away model choice per component (the thing a modular STT, LLM, TTS pipeline preserves), but it is one less integration to own.

About Inworld AI

FAQ

What happened to PlayHT?

Can I migrate my PlayHT voice clones to a new provider?

Should I reclone using AI-generated PlayHT audio?

How does Realtime TTS compare to PlayHT on quality?

What does Inworld Realtime TTS cost compared to PlayHT?

What is the easiest way to migrate code?