Speech to text and text to speech
There are two families of components for working with speech inside a flow:
- Speech to Text: they take an audio file, such as a voice note, a call recording or a meeting, and return what is said in it as text. They usually return one Data per file, with the text in the
textfield and the language inlanguage, ready for an agent to read, summarise or store. - Text to Speech: they take a text and return an audio file (MP3, WAV or OGG, depending on the component) to send over a messaging channel, attach to an email, store or play.
All of them can process a list of items with Batch Mode, and the new ones can also be used as a tool by an agent.
Where they are in the palette
In the A.I. section of the components panel, in two groups: Speech to Text and Text to Speech. Every component on this page is there, including those that used to appear in Media AI and AI Helpers.
Speech to text
| Component | Where it runs | Credential | Languages and notes |
|---|---|---|---|
| Parakeet Local | Local | None | 25 European languages, detected automatically. It does not recognise Catalan, Galician or Basque. |
| Whisper Local | Local | None | Detects the language, or you choose one of 12. |
| OpenAI Whisper | Cloud | OpenAI API | The language is chosen from 12; English by default. |
| Groq Whisper Speech to Text | Cloud | Groq API | More than 99 languages, detected automatically. Files up to 25 MB. |
| Deepgram Speech to Text | Cloud | Deepgram API | Detects the language; multi for audio that mixes languages. Files up to 2 GB. |
| Gemini Speech to Text | Cloud | Google Gemini | Any language, detected automatically. Files up to 2 GB. |
| AssemblyAI | Cloud | AssemblyAI key | The language is chosen from 12; English by default. |
| Cartesia Speech to Text | Cloud | Cartesia API | Does not detect the language: you set it, Spanish by default. |
| Inworld Speech to Text | Cloud | Inworld API | 30 languages, detected automatically. Files up to about 16 MB. |
| WitAI | Cloud | WitAi | The language of your Wit.ai app; returns the detected language. |
Text to speech
| Component | Where it runs | Credential | Languages and notes |
|---|---|---|---|
| Supertonic TTS | Local | None | 31 languages, detected sentence by sentence. WAV. |
| Kokoro TTS | Local | None | Spanish, English, French, Italian, Portuguese, German, Japanese, Korean and Chinese; you choose the language. WAV. |
| Piper TTS | Local | None | Each voice speaks one language. WAV. |
| OpenAI Text to Speech | Cloud | OpenAI API | Any language; the tone is described in words. MP3, WAV or OGG. |
| ElevenLabs | Cloud | Eleven Labs API | The voices in your account, custom ones included. MP3. |
| Gemini Text to Speech | Cloud | Google Gemini | Thirty voices, any language; the tone is described in words. WAV. |
| Cartesia Text to Speech | Cloud | Cartesia API | You set the language, Spanish by default. MP3 or WAV. |
| Deepgram Text to Speech | Cloud | Deepgram API | Spanish and English voices; the voice sets the language. MP3, WAV or OGG. |
| Inworld Text to Speech | Cloud | Inworld API | Inworld voices, your own included. MP3, WAV or OGG. |
How to choose
- Local or cloud. Local components run on NappAI’s servers: they need no key and no internet connection, and the audio or text never leaves the platform. Use them when privacy matters or when you have no account with any provider.
- In the cloud, the provider you already have. If you already use Deepgram, Groq, Cartesia, Inworld, Gemini or OpenAI in other components or in your voice agent, pick theirs: the key is filled in for you when you choose your credential, with nothing new to create.
- The language. If you don’t know which language the audio will be in, pick one that detects it. Cartesia Speech to Text does not.
- The audio format. For WhatsApp voice notes, a component that produces OGG or MP3; the local ones and Gemini produce WAV.
Measured in our test environment on 3 November 2026, with the sentence “Hola, buenos días. Su cita está confirmada para el martes a las diez y media.”:
| Provider | Text to speech | Speech to text |
|---|---|---|
| Inworld | 0.7 s (MP3) | 0.5 s |
| Deepgram | 1.1 s (MP3) | 0.4 s |
| Groq | — | 0.2 s |
| Cartesia | 0.7 s (WAV) | 0.2 s |
| OpenAI | 1.5 s (MP3) | — |
| Gemini | 3.3 s (WAV) | 4.5 s |
All five transcripts were exact. Deepgram, Groq, Cartesia and Gemini wrote the time as digits (“10 y media” or “10:30”); Inworld did not return the detected language.
They are not the Voice Agent nodes
The Voice Agent has its own speech-to-text (STT) and text-to-speech (TTS) nodes, such as Deepgram STT or Inworld TTS. Those nodes only connect to a Voice Agent and work live during a call, listening and speaking as it happens.
The components on this page are flow steps: they wait for the whole file or text and return a finished result. They cannot be connected to a Voice Agent.