Updated
AI models and media

Speech-to-Text & Text-to-Speech

Turn audio into text, or text into natural speech, for transcription, voice features and audio content.

For your framework or language:Next.jsAstro
Ask your AI about this, with this page as the source:ChatGPT ↗Claude ↗Perplexity ↗

Trends

Kokoro gained the most ground in package downloads: +105% against the median npm package over a year.

Public counts each month: pull requests on GitHub, package downloads, stars, Product Hunt followers. They show direction, not market share.

ToolDownloadsStarsProduct Hunt
OpenAI API37.4M/wk+60% vs npm—34.5k followers
Whisper2.1M/wk★ 110.4k—
Deepgram892k/wk+81% vs npm——
AssemblyAI513.3k/wk+38% vs npm——
ElevenLabs380.7k/wk18.9× vs npm*——
Kokoro145.7k/wk+105% vs npm★ 9.4k—
Cartesia104.7k/wk+80% vs npm——
Hume AIShutting down58.7k/wk+68% vs npm——
Fish Audio39k/wk★ 33k—

The real choice

Real-time or batch, and which direction. Streaming voice features need low latency above everything, while transcribing recordings is about accuracy and price per hour. Most providers lead in one direction — speech-to-text or text-to-speech — so many apps end up using two.

What makers chose223 live products, each with a source. Not market share.

Pick by situation

Tap the ones that are you — the tools that fit light up below.
  • How the voice sounds is part of the product — narration, characters, contentElevenLabsFish Audio
  • You're building live conversation where every millisecond of delay showsCartesiaDeepgram
  • Users upload recordings to be transcribed and analysedAssemblyAIDeepgram
  • Your product needs recordings and transcripts from Zoom, Meet or Teams callsRecall.ai
  • Audio must stay on your servers, or volume makes per-minute pricing add upWhisperKokoro
  • Voice is a side feature and one vendor matters more than fine controlOpenAI API

Compare the tools

Grouped by the side of the choice they answer, not ranked. Open a row for when to use it, the trade-off and what makers say.

ToolBest for
Voice generation3
ElevenLabsapi · self-hostableNatural, expressive voices and voice cloning for narration, content and characters.135+31 open source380.7k/wk18.9× vs npm*

Use it whenHow the voice sounds is part of the product.

Trade-offCredits run out fast on long audio, and cloning needs consent from the voice owner.

In Lovable, Replit

Used by AgentR, ai_licia, April, AskDeck, Automaticall and 130 more · in 31 open-source projects.

Transforming text into lifelike voices is a breeze with ElevenLabs' API. It adds the perfect touch of personality and professionalism to our videos. We couldn't do it without them!
Topview, the makerSep 2026 ↗
Cartesiaapi · self-hostableText-to-speech with very low time to first audio, for live conversation.29+6 open source104.7k/wk+80% vs npm

Use it whenYou're building a real-time voice agent where every millisecond of delay shows.

Trade-offBuilt for real-time use, with fewer creator tools such as dubbing and sound effects.

Used by Cekura, Coval, Harmony AI Planner, Martin, Mentara and 24 more · in 6 open-source projects.

Cartesia stood out for its ultra-low latency voice synthesis and natural-sounding speech. It helped us create more human-like real-time conversations while keeping the developer experience simple and flexible.
Thinnest AI, the makerSep 2026 ↗
Fish AudioapiText-to-speech and quick voice cloning with a large library of community voices in 30+ languages.9+6 open source39k/wk

Use it whenYou want many voice options, or to clone a voice from a short sample.

Trade-offCommunity voices vary in quality and rights; its open-weight models are under the Fish Audio Research License, so check its terms before self-hosting commercially.

Used by Calling Clones, InsForge, SUN, WaseiGo, CraftStory and 4 more · in 6 open-source projects.

I needed audio for about 5,000 Japanese dialogue lines. Most TTS makes Japanese sound like a train announcement. Fish Audio was the first one where my two characters sounded like actual people, and the price didn't scare me off.
WaseiGo, the makerSep 2026 ↗
Transcription2
Deepgramapi · self-hostableStreaming speech-to-text with low latency, for live calls and voice interfaces.65+12 open source892k/wk+81% vs npm

Use it whenYou transcribe audio as it happens and need words back in real time.

Trade-offIts text-to-speech is built for voice agents, with a set of prebuilt voices and no voice cloning.

Used by April, Breyta.ai, Cekura, Clipwing, CoAgentor and 60 more · in 12 open-source projects.

Deepgram gave us fast, reliable speech infrastructure and developer-first APIs that made it easy to build low-latency voice agents. Their startup program and support helped us move from prototype to production much faster.
Thinnest AI, the makerSep 2026 ↗
AssemblyAIapi · self-hostableAccurate transcription of recordings with speaker labels and extras like summaries and topic detection.26+11 open source513.3k/wk+38% vs npm

Use it whenUsers upload meetings, podcasts or calls to be transcribed and analyzed.

Trade-offFocused on speech-to-text; you need another vendor for speech output.

Used by Convo, IFTTT, InterviewFlowAI, Klariqo AI Voice Assistants, Kroto and 21 more · in 11 open-source projects.

AssemblyAI offers an advanced, easy-to-integrate speech recognition API that excels in transcribing audio data accurately. It supports a range of media types! Super handy to use.
Me.bot, the makerSep 2026 ↗
General LLM provider1
OpenAI APIapiAdding transcription and speech output on the same account you already use for text.137.4M/wk+60% vs npm

Use it whenVoice is a side feature and one vendor matters more than fine control.

Trade-offFewer voices and less tuning than specialist providers.

llms.txtIn Lovable, Replit, Rork

Used by Vercel.

The whisper-1 and tts-1 models provided the starting point for Orate.
Vercel, the makerSep 2026 ↗
Open model you run2
Whisperlibrary · open sourceTranscribing and translating audio in many languages with an open model on your own hardware.39 open source2.1M/wk

Use it whenAudio can't leave your servers, or you transcribe enough hours that per-minute pricing adds up.

Trade-offBatch-oriented, not built for real-time streaming; large models need a GPU for reasonable speed, and speaker labels need another tool.

No maker's product we track shows it for this yet · in 39 open-source projects.

Kokorolibrary · open sourceA small open-weight text-to-speech model that runs on a CPU or in the browser, under Apache 2.0.30 open source145.7k/wk+105% vs npm

Use it whenYou need speech output without per-character fees, or on-device.

Trade-offA fixed set of voices with no cloning, and less expressive than the leading hosted voices.

No maker's product we track shows it for this yet · in 30 open-source projects.

Meeting bots1
Recall.aiapiGetting recordings, transcripts and participant data from Zoom, Google Meet, Microsoft Teams and other calls through one bot API.21

Use it whenYour product is a notetaker, sales assistant or anything else that needs to be in users' meetings.

Trade-offBilled per recording hour on top of any transcription, and the bot is visible to everyone in the call.

Used by Ariso, Brev.io, Cuckoo, Katalyst, Mina and 16 more.

Katalyst joins calls across every platform a sales team uses. Recall let us do that without building four fragile integrations ourselves.
Katalyst, the makerSep 2026 ↗

Before you choose

How to approach it

Decide whether audio is live (calls, voice UI) or recorded (uploads, meetings); the right provider differs. Test with your own audio — accents, background noise and domain terms change accuracy more than benchmarks show. For TTS, listen to full paragraphs, not single sentences, before choosing a voice.

Common mistakes
  • Choosing a speech-to-text provider on clean benchmark audio, then discovering it struggles with your users' accents, jargon and background noise.
  • Cloning a voice without the speaker's documented consent, which breaks provider terms and, in many places, the law.
  • Sending whole recordings to a streaming API or live audio to a batch one, paying for latency you don't need or waiting on latency you can't accept.

Other options

Real choices most makers here won't need to weigh.

Decided alongside

What the 223 makers' products here chose for their other decisions.