Oct 2026
Emotion is a direction, not a dataset
How Forge v3 learned to take stage directions — [whispers], [laughs], [sighs] — from text alone, and why we evaluate acting, not just pronunciation.
Giving every idea
a voice of its own
One platform for the people and machines that speak. ForgeStudio for creators, ForgeAgents for customer experience, and the voice API developers build on — all from the same research lab.
Trusted by developers, studios and support teams
Read all storiesnorthpeak
Helios Media
cartwheel
stripeline
MOSAIC
ferrous&co
KODA GAMES
Brightline
atlas telecom
LUMEN
vantage
Pinecrest
orbita
FIELDNOTE
Casa Audio
nexbank
trove
WELLSPRING
Two platforms, one research foundation
Every model we train ships to both sides of the house — the studio where content is made, and the agents that hold real conversations.
ForgeStudio
Generate lifelike speech, dub video, compose music and design sound — in one editor.
Project · Harbour Lights, ch. 12Rendered
Voice · Aria — NarrationForge v3
Dubbing · ES, DE, JA3 / 3 done
Exports this month1,284
ForgeAgents
Configure, deploy and monitor conversational agents across phone, chat and messaging.
Agent · Support — BillingLive
Calls today7,412
Avg. handle time2:48
Escalations to humans6.1%
Write a script once and hear it performed in any voice and more than forty languages. Cut films, ads, audiobooks, podcasts and social clips without ever booking a booth.
Beyond the last buoy, the water changed colour — a deep green that the charts refused to name, as though the sea kept one shade for itself.
The lighthouse keeper logged it without comment, the way he logged everything the island did not explain.
Chapter twelve · Harbour Lights
Aria — Narration · English · Forge v3
On the first morning of the voyage, the crew found frost on the inside of the sails. [whispers] Nobody mentioned it at breakfast. [laughs] By noon, the whole ship was pretending not to be cold.
Emotion tags perform as they’re read
All-in-one AI editor
Podcasts, audiobooks and voiceovers, edited on a timeline built on our full audio research stack.
Ultra-realistic speech
Controllable, expressive speech with breath, pace and emphasis — layered across 40+ languages.
Music
Studio-length tracks in any genre, with vocals or purely instrumental.
SFX
Custom sound effects, soundscapes and ambience from a sentence of description.
Voices
Clone your own voice, design one from a prompt, or license from 6,000+ library voices.
Dubbing & Video
Translate and re-voice video while keeping the original speaker’s tone and timing.
Voice Library
Six thousand voices. Start with these nine.
Every voice is licensed for commercial use and rendered by the same models as the studio. Press play — each sample is spoken live by your browser, tuned to the voice’s pitch and pace.
No voices match that search. Try a different word — “deep”, “bright”, “elder”.
Configure, deploy and monitor natural, human-sounding agents in 40+ languages, with low-latency voice or chat and a full audit trail of every decision.
Can I get a refund for order 4521?
Of course. Can you confirm the email on the order?
It’s under my work email.
Verified. I’ve started the refund — it lands in 3–5 business days.
$ Refund issued · #4521
Omnichannel agents
One agent listens, reads and replies across phone, chat, email and messaging apps.
Resolution rate · last 7 days
86.2%
ForgeAgentsLegacy IVR
Analytics — measure resolution, sentiment and cost per conversation, then replay any call.
Testing
Simulate thousands of callers — accents, interruptions, edge cases — before a single real customer hears the agent.
Guardrails
Hard rules for what an agent may promise, refund or escalate — enforced on every turn, logged for compliance.
Workflows
Branching flows that check stock, book appointments and update your CRM mid-conversation, in real time.
cartwheel
Resolving delivery questions by voice while riders keep both hands on the road.
Get startedForgeAPI
Or build anything with a full shelf of APIs
Text to Speech API
Independently benchmarked among the most lifelike speech models available. Pick a model per job — consistency, latency or emotional range — and switch with one parameter.
Forge Flash~82 ms latency, built for live conversation
Forge MultilingualSteadiest delivery across 40+ languages
Forge v3Our most expressive model — tags for tone, pace and emotion
Forge DubRe-voices video in the original speaker’s voice
Speech to Text API
ForgeScribe transcribes calls, meetings and media with speaker labels and word-level timestamps — including the ums, overlaps and crosstalk other engines clean away.
Forge Scribe98.1% word accuracy on conversational audio
DiarizationUp to 32 speakers, labelled automatically
# transcribe a support call
result = vf.transcribe(
file="call_88412.mp3",
diarize=True,
)
for seg in result.segments:
print(seg.speaker, seg.start, seg.text)
Music & Sound API
Compose licensed, studio-grade tracks from a text brief, or generate one-shot effects and beds that loop cleanly under dialogue.
Forge MusicFull arrangements up to 5 minutes, stems included
Forge SFXEffects rendered to your exact duration
const track = await vf.music.compose({
brief: "Slow piano and felt strings, hopeful, no drums",
durationSeconds: 90,
stems: true,
});
Pricing
Start free. Scale when the audience does.
Credits cover speech, dubbing and music. Agent minutes and transcription are metered separately, always shown before you spend them.
How much speech do you make?
Drag to your monthly volume and we’ll point at the plan that covers it cheapest. 120,000 characters / month
Best fit
Creator · $22/mo
Covers 120,000 of your 120,000 characters.
What teams make once everything can speak
Research that redefines how people and machines talk
We train our own foundation models — starting with a voice model that could hesitate, laugh and whisper, and now reaching into transcription, music and agents that act.
Forge v3
Our fastest and most expressive voice model yet — directed mid-sentence with plain-language tags.
Jun 2026
Sep 2026
Latency budgets for live conversation
Humans notice a pause at 300 milliseconds. The engineering diary of getting streamed speech under a tenth of that, end to end.
Aug 2026
Transcription that keeps the mess
Overlaps, false starts and crosstalk carry meaning. Forge Scribe preserves them — and support QA teams finally get the full call.
FAQ
Questions, answered out loud
The short version of everything teams ask before their first render, call or API key.
Yes. Record three minutes of clean audio, verify it is you with a spoken captcha, and your clone is ready in about a minute. Clones are locked to your account, watermarked in the audio stream, and can be deleted permanently at any time.
Speech generation covers 40+ languages including English, Indonesian, Spanish, German, Japanese, Hindi and Arabic. The same voice can read across languages while keeping its timbre, and dubbing preserves the original speaker in every target language.
One credit is roughly one character of generated speech on Forge Multilingual, with Forge v3 priced slightly higher per character and Flash slightly lower. Transcription is billed per minute of audio and agent conversations per minute connected. The usage page shows every charge against the render or call that caused it.
They act through tools you approve: checking orders, issuing refunds within limits you set, booking and rescheduling, updating CRM records. Anything outside its guardrails is escalated to a human with a full transcript and summary attached.
Every render carries an inaudible provenance watermark and C2PA metadata identifying it as synthetic. Voice cloning requires proof of consent, and a public blocklist prevents cloning of people who have opted out.
Every account starts with 10,000 free credits a month and full API access — no card required. The same keys work across speech, transcription, music and agents, with per-key spend limits for production safety.
Your next project already
knows what it sounds like.
Create a free account and make your first render in under a minute — 10,000 credits, every model, no card.