Speech AI, voice AI and conversational AI are overlapping labels, not separate technologies. Speech AI names the component technologies that handle spoken language, such as recognition and synthesis. Conversational AI names systems that hold a dialogue, whether typed or spoken. Voice AI is the product-level term for machines you talk to or that talk to you, and it draws on both.
Key takeaways
- Speech AI is defined by its input or output: audio of human speech. It includes transcription tools that have no conversation at all.
- Conversational AI is defined by behaviour: multi-turn dialogue. A text chatbot is conversational AI with no speech involved.
- Voice AI is where the two meet in most products: a conversation carried over audio. The term is also stretched to cover one-way speech tools.
- The labels are marketing categories as much as technical ones, so read what a vendor actually provides rather than the category it claims.
The three terms#
Speech AI#
Speech AI, sometimes "speech technology" or "spoken language processing", covers models that take speech in or produce speech out. The core members are automatic speech recognition, text-to-speech, speaker recognition, diarisation, spoken language identification and keyword spotting. The field has its own conferences, such as Interspeech and ICASSP, and its own metrics, starting with word error rate.
Speech AI is component technology. A speech recogniser does not know it is in a conversation; it converts audio to text and stops there.
Conversational AI#
Conversational AI covers systems that exchange multiple turns with a person and keep track of context between them. Its lineage runs through natural language understanding, dialogue state tracking and dialogue management, and today mostly through large language models. The channel is incidental: the same assistant can run in a chat window, a messaging app or on a phone line.
When the channel is text, conversational AI involves no speech technology at all. When the channel is voice, it needs speech AI at both ends.
Voice AI#
Voice AI is the newest and loosest of the three. Its narrow meaning is a conversational system carried over audio: a voice agent, an in-car assistant, a phone bot. Its broad meaning, common in industry reports and funding announcements, is any AI product in which voice is central, which includes dictation, dubbing, voice cloning and call analytics.
Side by side#
| Speech AI | Conversational AI | Voice AI | |
|---|---|---|---|
| Defined by | Modality (spoken audio) | Behaviour (dialogue) | Product experience (talking to a machine) |
| Typical output | Text, audio, labels, timestamps | A reply or an action | A spoken reply or a transcript |
| Example without the others | Batch transcription of court hearings | A support chatbot on a website | Rare: it usually implies at least speech AI |
| Key metrics | WER, RTF, MOS for synthesis | Task success, containment, user satisfaction | End-to-end latency, task success |
| Research home | Interspeech, ICASSP | ACL, SIGDIAL | Spread across both |
Where they meet: the voice agent#
A phone agent that books appointments is all three at once. It uses speech AI to transcribe the caller and synthesise replies, conversational AI to track what has been agreed and decide the next question, and it is sold as voice AI. The voice agent architecture article shows how the parts connect.
The overlap is also where the distinctions matter most in practice. A conversational AI team used to text will find that voice adds problems that do not exist in chat:
- Recognition errors. Text users type what they mean. Voice users are transcribed, and a caller who says I'd like to cancel may arrive as I'd like to counsel.
- Turn boundaries. In chat, pressing enter ends a turn. In speech, the system must infer it from pauses, intonation and wording. See end-of-turn detection.
- Interruptions. Callers talk over the system. See turn-taking and barge-in.
- Time pressure. A chat reply that takes three seconds is fine. A spoken reply that takes three seconds is a silence on the line.
- Formatting. Numbers, dates and codes are spoken in words and must be converted, see inverse text normalisation, and replies must be written for the ear, without bullet points or URLs.
How the terms evolved#
"Speech recognition" and "speech synthesis" are decades old; the umbrella "speech technology" was standard in academic and industrial research long before "AI" became a common prefix. "Conversational AI" spread in the late 2010s with chatbot platforms and smart-speaker assistants, when dialogue systems were built from intent classifiers and slot fillers. "Voice AI" became prominent in the 2020s, particularly after large language models made open-ended spoken conversation practical and a wave of voice agent products appeared.
Older terms still appear in procurement and standards documents: interactive voice response (IVR) for menu-driven phone systems, and spoken dialogue system in the research literature. An IVR with speech input is an early form of what is now called a voice agent.
Which term to use#
For precision, name the component. If a system transcribes, say speech recognition. If it holds a dialogue, say conversational system or dialogue system. If it does both over audio, voice agent is the clearest current term. Reserve voice AI for the category, not for a specific capability.
When evaluating vendors, the category name tells you little. A "voice AI platform" might provide only orchestration and resell recognition and synthesis from others, or it might train its own models. Ask which layers of the stack are built in-house. See how to evaluate a speech-to-text vendor.
Common questions#
Is a chatbot conversational AI?
Yes, if it holds a multi-turn dialogue. It is not speech AI or voice AI unless it accepts or produces speech.
Is transcription voice AI?
In the broad industry sense, yes. Technically it is speech AI: there is no dialogue, and the output is text.
Where does audio intelligence fit?
Audio intelligence, such as summarisation, sentiment and topic detection on calls, is usually grouped with speech AI because it starts from a transcript of recorded speech. It often uses the same language models as conversational AI.
Is a speech-to-speech model speech AI or conversational AI?
Both. It handles audio directly, which is speech AI, and it holds a dialogue, which is conversational AI. See speech-to-speech models vs pipelines.
References#
- Jurafsky, D., & Martin, J. H. Speech and Language Processing (3rd ed. draft), chapters on chatbots, dialogue systems and speech recognition. Stanford.
- Wilpon, J. G., Rabiner, L. R., Lee, C.-H., & Goldman, E. R. (1990). Automatic recognition of keywords in unconstrained speech using hidden Markov models. IEEE Transactions on Acoustics, Speech, and Signal Processing, 38(11).
- Young, S., Gašić, M., Thomson, B., & Williams, J. D. (2013). POMDP-based statistical spoken dialog systems: a review. Proceedings of the IEEE, 101(5).
- McTear, M. (2020). Conversational AI: Dialogue Systems, Conversational Agents, and Chatbots. Morgan & Claypool.