In short: This AI voice glossary defines 47 common terms used in text to speech, voice cloning, speech recognition and digital audio, from accent and bitrate to zero-shot. Each entry is one to three sentences in plain English, written to be vendor-neutral, so creators, teachers and developers can understand AI voice tools quickly.
What is an AI voice glossary for?
AI voice tools borrow words from linguistics, audio engineering and machine learning. This glossary translates that jargon into everyday language so you can choose settings with confidence, understand tool documentation, and explain your work to clients or students. Terms are grouped alphabetically; related terms are cross-referenced in the definitions.
New to the topic? Start with our explainers what is text to speech, what is voice cloning and what is AI dubbing.
AI voice terms A–L
| Term | Definition |
|---|---|
| Accent | The way a language is pronounced in a particular region or community, such as American, British, Australian or Indian English. TTS voices are usually labeled by language and locale to reflect accent. |
| Acoustic model | The part of a speech system that links text or phonemes to sound. In TTS it predicts what the audio should look like; in speech recognition it maps sounds to likely phonemes. |
| AI voice | Any synthetic voice produced by artificial intelligence, including text-to-speech voices, cloned voices and designed voices. |
| ASR (automatic speech recognition) | Technology that converts spoken audio into written text. Also called speech to text. |
| Bitrate | How much data is used per second of compressed audio, measured in kilobits per second (kbps). For MP3, 128 kbps is fine for speech, while 192–320 kbps preserves more detail for music. |
| Breath sounds | The small inhalations between phrases. Natural-sounding neural voices often include subtle breaths, which make long narration feel more human. |
| Character limit | The maximum amount of text a TTS tool converts in one go. Long scripts are split into parts and joined afterward. |
| Concatenative synthesis | An older TTS method that stitches together small recorded pieces of a real voice. It can sound natural but often has audible joins. |
| Consent | Clear permission from a person to record, clone or use their voice. Ethical voice cloning starts with documented consent. |
| Deepfake audio | Synthetic audio designed to deceive listeners into believing a real person said something. Distinct from consented, disclosed uses of voice cloning. |
| Diarization | Short for speaker diarization: automatically working out “who spoke when” in a recording with several people, and labeling each segment by speaker. |
| Dubbing | Replacing the spoken language of a video with a new voice track in another language. AI dubbing automates translation and re-voicing. |
| Emphasis | Extra stress placed on a word to change meaning (“I didn’t say that”). In scripts, short sentences and punctuation help TTS place emphasis well. |
| End-to-end model | A neural network that goes from text to audio (or audio to text) in one learned system, instead of separate hand-built stages. |
| Fine-tuning | Further training an existing AI model on a smaller, specific dataset, such as many recordings of one voice, to specialize it. |
| Formant | A resonant frequency of the human vocal tract that shapes vowel sounds. Formant synthesis, used in early TTS, generated speech by modeling these resonances. |
| Grapheme-to-phoneme (G2P) | The step that converts written letters (graphemes) into speech sounds (phonemes), for example deciding how to pronounce “tough,” “though” and “through.” |
| Hallucination (transcription) | When a speech recognition system outputs words that were not actually spoken, often during silence or noise. Proofreading catches it. |
| Intonation | The rise and fall of pitch across a sentence. Questions in English usually rise at the end; statements fall. |
| IVR (interactive voice response) | Automated phone menus (“Press 1 for billing”) that use recorded or synthesized voice prompts. See phone greetings and IVR. |
| Latency | The delay between sending text or audio and receiving the result. Low latency matters for live assistants and phone bots. |
| Lip-sync | Matching mouth movements in video to the audio. In AI dubbing, lip-sync is a separate step that edits the picture. |
| Locale | A language plus region code, such as en-US (English, United States) or es-MX (Spanish, Mexico), used to pick voices with the right accent and conventions. |
| Loudness (LUFS) | A measure of perceived loudness, in Loudness Units relative to Full Scale. Streaming platforms normalize audio to a target loudness, so consistent LUFS avoids volume jumps. |
AI voice terms M–Z
| Term | Definition |
|---|---|
| Mel spectrogram | A picture of sound showing how energy across frequencies changes over time, scaled to match human hearing. Many neural TTS systems predict one before generating audio. |
| MP3 | The most widely supported compressed audio format. Ideal for sharing voiceovers and podcasts because files are small and play everywhere. |
| Multilingual voice | A single TTS voice that can speak more than one language while keeping the same vocal identity. |
| Neural TTS | Text to speech generated by deep neural networks trained on recorded speech. The current standard for natural-sounding voices. |
| Normalization (text) | Rewriting text into its spoken form before synthesis: “$5.99” becomes “five dollars and ninety-nine cents,” “St.” becomes “Street” or “Saint” depending on context. |
| Phoneme | The smallest unit of sound that distinguishes meaning in a language, such as the /b/ and /p/ that separate “bat” from “pat.” American English has roughly 40 phonemes. |
| Pitch | How high or low a voice sounds, set by the frequency of vocal fold vibration. Most TTS tools let you raise or lower pitch slightly. |
| Prosody | The rhythm, stress, pitch and timing of speech: the “music” of how something is said. Good prosody is the biggest difference between natural and robotic TTS. |
| Sample rate | How many times per second audio is measured, in hertz (Hz). 44.1 kHz is CD quality; 22.05–24 kHz is common for speech; 16 kHz is typical for speech recognition and phone-quality input. |
| Sibilance | Harsh “s,” “sh” and “z” sounds. A de-esser reduces sibilance in voiceovers and isolated vocals. |
| Source separation (stem separation) | Splitting a mixed recording into parts such as vocals, drums and bass. Vocal removers and acapella extractors use it. |
| Speaker embedding | A compact numerical fingerprint of a voice’s characteristics, used by voice cloning systems to reproduce that voice. |
| Speech-to-speech (voice conversion) | Transforming a recording of one voice into another voice while keeping the original words and timing. Also called a voice changer. |
| SSML (Speech Synthesis Markup Language) | A W3C standard XML format for controlling TTS output with tags for pauses, emphasis, pronunciation, pitch and rate. Many simple tools use punctuation and sliders instead. |
| Stability / variation | How consistent a generated voice is from one take to the next. Higher variation can sound more expressive but less predictable. |
| Timbre | The tone color that makes two voices at the same pitch sound different, for example warm, bright, breathy or gravelly. |
| Timestamps | Time markers in a transcript showing when each word or sentence was spoken. Used for captions, subtitles and editing. |
| Vocoder | The component that turns a predicted acoustic representation, such as a mel spectrogram, into an audible waveform. |
| Voice cloning | Creating a synthetic copy of a specific person’s voice from a recording, so it can speak new text. See what is voice cloning. |
| Voice design | Creating a new synthetic voice from a written description, such as “warm, middle-aged female, Southern US accent,” rather than from a recording. |
| WAV | An uncompressed audio format with no quality loss. Larger than MP3; preferred for editing and mastering. |
| WER (word error rate) | The standard accuracy measure for speech recognition: the percentage of words substituted, deleted or inserted compared with a correct transcript. Lower is better. |
| Words per minute (WPM) | Speaking speed. Conversational American English narration typically falls around 140–170 WPM. |
| Zero-shot | An AI model’s ability to handle something it was not specifically trained on, such as cloning a new voice from a few seconds of audio without retraining. |
How do I put these terms into practice?
- Hear prosody in actionPaste the same sentence into text to speech twice, once with commas and once without, and compare the rhythm.
- Try a different localePlay the same line in en-US, en-GB and en-AU voices to hear how accent changes vowels and intonation.
- Test zero-shot cloningClone your own voice from a 5–30 second sample with voice cloning.
- Check timestampsTranscribe a short clip with sentence timestamps in audio to text to see how captions are timed.
- Compare formatsGenerate a track in the music generator, download it as MP3 and as WAV (the arrow next to Download), and listen to both on good headphones.
Building with AI voices?
Developers who want to put these voices inside an app, website or workflow can use the Speechma API (plans from $9/month, 1,000+ voices, instant key). For long-form narration in the browser, SpeechmaPRO.com converts up to 10,000 characters at once.
Frequently asked questions
What is the difference between TTS and ASR?
TTS (text to speech) turns text into audio. ASR (automatic speech recognition) turns audio into text. They are opposite directions of the same field.
What is the difference between prosody and intonation?
Intonation is the pitch pattern of a sentence. Prosody is broader: it includes intonation plus rhythm, stress, pauses and speed.
What sample rate should I use for voiceovers?
For video, 44.1 kHz or 48 kHz is standard. For speech-only uses like podcasts, 44.1 kHz is common. Speech recognition works well with 16 kHz audio.
What MP3 bitrate is good enough for speech?
128 kbps is plenty for spoken voice. Use 192–320 kbps for music or when you plan to edit the file further.
What is a good word error rate?
Lower is better. Clean, single-speaker English audio often transcribes with a low WER, while noisy, accented or overlapping speech produces higher error rates.
Do I need to learn SSML to use AI voices?
No. Many tools, including Speechma, use simple controls instead. In Speechma you add a pause with a short tag like [break 2s] (or a pause button), and adjust pitch, speed and volume with sliders.
The Speechma newsletter
Make better voiceovers, faster
- New voices and features the day they launch
- Script templates and tricks that make AI voices sound human
- Creator stories, community picks and early access
