If you're building an app that needs to listen to people talk or talk back to them — a phone support bot, a dictation feature, a voice assistant — you're going to need a "voice AI" service under the hood. These tools handle the hard parts: turning speech into text (transcription), turning text into speech (synthesis), or both at once wired into a live phone call. Some run in the cloud and charge per minute; a couple run entirely on the user's device for free. Below are seven real options, each solving a slightly different piece of the voice puzzle.
What actually differs between these tools
Three flavors show up in this space. Speech-to-text (STT) and text-to-speech (TTS) API providers (Deepgram, Cartesia, Rime) sell you the raw building blocks — you still wire them into your own app logic. Voice agent platforms (Vapi, Retell AI) bundle STT, an LLM, TTS and phone connectivity into a ready-to-configure phone bot. On-device toolkits (KittenTTS, Moonshine) skip the cloud API entirely and run the model locally — free, private, but limited by whatever hardware it's running on.
Deepgram — one bill for both transcription and speech generation
Deepgram is a speech recognition and voice-generation API that lets developers add accurate, fast transcription and natural-sounding speech to any app, both under one account. It's aimed at developers building voice AI applications and enterprises with custom model needs.
Pricing: $200 free credit to start. Speech-to-text runs roughly $0.0048-$0.0065/minute, text-to-speech about $0.015-$0.030 per 1,000 characters, and the bundled Voice Agent API (STT+LLM+TTS together) around $0.075/minute. Paid plans start near $4K/year, enterprise is custom.
Strengths: both STT and TTS on one API and one invoice; a bundled Voice Agent API that handles the full pipeline; built-in audio intelligence like summarization, sentiment and speaker diarization.
Limits: it isn't the clear latency or expressiveness leader against specialists like Cartesia or Rime, the pricing is competitive but not a standout discount, and the Voice Agent API's per-minute cost adds up fast at high call volume.
Verdict: pick Deepgram when you want one vendor covering both directions of voice instead of stitching together two separate APIs.
Cartesia — built for near-instant, natural-sounding speech
Cartesia builds real-time voice AI models designed to be the audio layer behind voice agents that need to feel like an actual conversation, not a robotic response with a lag.
Pricing: a free tier with 20,000 credits/month, then Pro at $5/month for 100K credits, Startup at $49/month for 1.25M credits, and Scale at $299/month for 8M credits. Voice agent calls run about $0.06/minute, telephony about $0.014/minute.
Strengths: a State Space Model architecture built specifically for low-latency real-time speech, deployment options across cloud, on-premise and on-device, and a very cheap $5/month tier to test before committing.
Limits: latency and quality leaderboard claims are self-reported and worth verifying independently, the feature set is narrower than bundled platforms like Deepgram's Voice Agent API, and per-minute costs still add up at high call volume.
Verdict: pick Cartesia when raw conversational latency is the thing you're optimizing for, especially in regulated industries needing on-premise deployment.
Rime — text-to-speech tuned for real conversation, not audiobooks
Rime is a text-to-speech API built by linguists, offering hundreds of natural-sounding voices across 50+ languages and tuned specifically for back-and-forth conversation rather than narration.
Pricing: $0.05 per 1,000 characters, with 3,000 free minutes for new accounts and custom volume pricing for enterprise.
Strengths: 600+ voices across 50+ languages and regional dialects, deterministic pronunciation controls for proper nouns and technical terms, and sub-100ms latency with on-premise/VPC deployment for regulated industries.
Limits: it's TTS-focused only — no built-in speech recognition like Deepgram offers — latency claims are self-reported like everyone else's in this space, and enterprise volume pricing requires a sales call to pin down.
Verdict: pick Rime when voice naturalness and language/dialect coverage matter more than getting a full STT+TTS bundle from one vendor.
Vapi — orchestrates a full phone-answering voice agent
Vapi is a developer platform that wires together speech recognition, an AI model and speech generation into a working phone-answering voice agent, aimed at SaaS developers and enterprises building phone-based AI agents.
Pricing: Build tier runs about $0.05/minute in hosting plus $10 per concurrent line/month, with model costs passed through or waived if you bring your own API key. Scale tier is an annual contract with custom per-minute rates.
Strengths: orchestrates STT, LLM, TTS and PSTN telephony into one working pipeline, targets sub-500ms latency for natural-feeling calls, and lets you bring your own API key to control underlying model cost and quality.
Limits: it competes closely with Retell AI and Bland AI with limited fundamental differentiation between them, the per-concurrent-line pricing plus usage adds real cost at scale, and underlying model provider costs are a separate line item to budget for.
Verdict: pick Vapi when you want developer-level control over the voice agent pipeline rather than a no-code drag-and-drop builder.
Retell AI — no-code phone agents for support and sales teams
Retell AI is a no-code-friendly platform for building AI phone agents that handle customer service and sales calls, with drag-and-drop configuration instead of writing pipeline code.
Pricing: pay-as-you-go from $0.07 to $0.31/minute depending on configuration, with free credits to start and custom enterprise pricing.
Strengths: drag-and-drop agent configuration that needs no code for most changes, a roughly 600ms latency target for natural conversation flow, and post-call AI quality assurance that automates call review.
Limits: it's a crowded category — Vapi, Bland, Synthflow, Thoughtly, Decagon and Sierra all overlap here — the wide pricing range makes real cost hard to predict without testing, and there's less developer-level control over the underlying pipeline than code-first competitors offer.
Verdict: pick Retell AI when your team wants to configure a phone agent visually without writing integration code.
KittenTTS — a tiny, free, open-source voice that runs without a GPU
KittenTTS is a tiny, open-source text-to-speech library — as small as 25MB — that produces natural-sounding voices on ordinary CPUs, no GPU required and no per-character bill.
Pricing: 100% free and open source (Apache 2.0), with optional paid commercial support and custom voices.
Strengths: a minuscule model (25-80MB) that runs without a GPU, completely free with no per-character cost, 8 built-in voices at 24kHz output, and built-in text normalization for dates and abbreviations.
Limits: fewer voices and less variety than paid cloud services, voice quality trails the best commercial AI voice clones on the market, and commercial support is a paid add-on only.
Verdict: pick KittenTTS when you need speech synthesis embedded in a mobile or edge app with zero per-call cost and no GPU available.
Moonshine — an on-device toolkit for STT, TTS and intent recognition
Moonshine is an open-source toolkit for building voice AI that runs fully on-device, with no internet connection required and no per-request API bill.
Pricing: 100% free and open source (MIT license), with models downloadable via CLI or auto-downloaded by mobile SDKs, no cost per call.
Strengths: runs 100% on-device with zero network dependency, claims better accuracy than Whisper Large V3 at a much smaller size, native bindings for iOS/Android/desktop/Raspberry Pi, and bundles STT, TTS and intent recognition in one toolkit.
Limits: performance is capped by the target device's compute power, it's less plug-and-play than a simple cloud API call, and the ecosystem/community is smaller than Whisper's.
Verdict: pick Moonshine when you're building an embedded voice assistant or offline transcription feature and want to avoid API dependency entirely.
Quick comparison
| Tool | Type | Best for | Starting price |
|---|---|---|---|
| Deepgram | STT + TTS API | One vendor for both directions of voice | $200 free credit, then usage-based |
| Cartesia | Real-time voice models | Lowest-latency conversational speech | Free tier, $5/mo Pro |
| Rime | TTS API | Voice naturalness across 50+ languages | $0.05/1K characters |
| Vapi | Voice agent platform | Developer-controlled phone agents | ~$0.05/min + $10/line/mo |
| Retell AI | Voice agent platform | No-code phone agent builder | $0.07-$0.31/min |
| KittenTTS | On-device TTS | Free, GPU-free embedded synthesis | Free (open source) |
| Moonshine | On-device STT/TTS toolkit | Fully offline voice AI | Free (open source) |
There's no single "best" voice AI tool because these products solve different problems: raw API building blocks versus ready-made phone agents versus free on-device models. If you're prototyping a phone bot, start with Vapi or Retell AI to get something working fast. If you're already deep in a custom pipeline and need best-in-class transcription or synthesis, Deepgram, Cartesia and Rime are the specialists to benchmark against each other. And if cost or privacy rules out sending audio to the cloud at all, KittenTTS and Moonshine are the only two on this list that don't require a network connection.