Human-Like Conversational
Text-to-Speech from Small Voice Models

Human-Like Conversational
Text-to-Speech from
Small Voice Models

Human-Like Conversational
Text-to-Speech from Small Voice Models

Written by Sohaib Ahmad

Written by Sohaib Ahmad

Written by Sohaib Ahmad

Published on 18/11/2026

Published on 18/11/2026

Published on 18/11/2026

Abstract green waveform made of glowing dots, representing synthesised speech.
Abstract green waveform made of glowing dots, representing synthesised speech.

Natural speech from a small model, running on a CPU.

Natural speech from a small model, running on a CPU.

Natural speech from a small model, running on a CPU.

Teams putting voice into glasses, robots and phones keep reaching the same fork. Cloud text to speech sounds human, but it needs a network, sends data off the device and bills by the minute. Local models run offline, but they sound robotic or blow the memory and power budget.


Human-like conversational text to speech from a small, custom model removes that choice when the model is designed for the device’s CPU from the first experiment. NeuTTS-2M, for example, speaks seven languages from 60M active parameters and runs in real time on two CPU threads.


This guide will cover what makes conversational speech sound human, why conversation is harder than narration, how to measure quality honestly, and what to check before you ship. We build these models at Neuphonic, so wherever we quote our own numbers we say how we measured them.

Teams putting voice into glasses, robots and phones keep reaching the same fork. Cloud text to speech sounds human, but it needs a network, sends data off the device and bills by the minute. Local models run offline, but they sound robotic or blow the memory and power budget.


Human-like conversational text to speech from a small, custom model removes that choice when the model is designed for the device’s CPU from the first experiment. NeuTTS-2M, for example, speaks seven languages from 60M active parameters and runs in real time on two CPU threads.


This guide will cover what makes conversational speech sound human, why conversation is harder than narration, how to measure quality honestly, and what to check before you ship. We build these models at Neuphonic, so wherever we quote our own numbers we say how we measured them.

What makes conversational text to speech sound human?

What makes conversational text to speech sound human?

A synthetic voice sounds human when its prosody fits the meaning, its pauses fall where a speaker would breathe or think, it says names and numbers correctly, and it never skips or invents a word. Get one of those wrong and listeners notice within a sentence.

A synthetic voice sounds human when its prosody fits the meaning, its pauses fall where a speaker would breathe or think, it says names and numbers correctly, and it never skips or invents a word. Get one of those wrong and listeners notice within a sentence.

Prosody: pitch, rhythm, stress and pacing

Prosody: pitch, rhythm, stress and pacing

Prosody is the melody and timing of speech, the pitch contour (F0), syllable length, stress and speaking rate. Stress carries meaning. “I said Tuesday” defends a statement; “I said Tuesday” corrects a date.

Models lose this easily. In a study of phoneme-level prosody prediction (arXiv:2211.01327, 2022), the model listeners rated most natural had flat prosody, and adding variation lowered MOS and raised word error rate. Standard metrics reward a safe, even delivery. In conversation that delivery sounds boring, because real speakers stress the word that carries new information.

Prosody is the melody and timing of speech, the pitch contour (F0), syllable length, stress and speaking rate. Stress carries meaning. “I said Tuesday” defends a statement; “I said Tuesday” corrects a date.

Models lose this easily. In a study of phoneme-level prosody prediction (arXiv:2211.01327, 2022), the model listeners rated most natural had flat prosody, and adding variation lowered MOS and raised word error rate. Standard metrics reward a safe, even delivery. In conversation that delivery sounds boring, because real speakers stress the word that carries new information.

Intonation that follows the meaning of each turn

Intonation that follows the meaning of each turn

Yes and no questions in English usually rise at the end (“Did you lock the door?”) and statements fall. A flat F0 contour across both is the most recognisable robotic marker.

A “question mark means rise” rule fails too. Wh-questions such as “Where did you park?” normally fall. The model has to read what the sentence is doing, which is why language-model-based TTS handles intonation better than rule-driven pipelines, because it reads the whole sentence before predicting the first sound.

Yes and no questions in English usually rise at the end (“Did you lock the door?”) and statements fall. A flat F0 contour across both is the most recognisable robotic marker.

A “question mark means rise” rule fails too. Wh-questions such as “Where did you park?” normally fall. The model has to read what the sentence is doing, which is why language-model-based TTS handles intonation better than rule-driven pipelines, because it reads the whole sentence before predicting the first sound.

Pauses and phrasing that come from context, not punctuation

Pauses and phrasing that come from context, not punctuation

Punctuation driven TTS pauses at every comma. People pause before a hard word, after a clause that sets up a point, and hardly at all inside a familiar list. Punctuation still helps, and our NeuTTS guidance tells developers that commas, full stops and question marks guide rhythm. The model treats them as hints and uses context to decide how long each pause should be.

Punctuation driven TTS pauses at every comma. People pause before a hard word, after a clause that sets up a point, and hardly at all inside a familiar list. Punctuation still helps, and our NeuTTS guidance tells developers that commas, full stops and question marks guide rhythm. The model treats them as hints and uses context to decide how long each pause should be.

Saying names, numbers and jargon correctly

Saying names, numbers and jargon correctly

One mispronounced customer name undoes a lot of good prosody. Conversational products read “£4.50”, “14:30”, “SKU 7Q-22” and drug names that narration rarely meets. Context-aware formatting reads numbers, dates, units and codes the way a person would. Pronunciation control fixes a word with a rule, through a pronunciation dictionary or International Phonetic Alphabet (IPA) tags, so you never retrain to correct a brand name.

One mispronounced customer name undoes a lot of good prosody. Conversational products read “£4.50”, “14:30”, “SKU 7Q-22” and drug names that narration rarely meets. Context-aware formatting reads numbers, dates, units and codes the way a person would. Pronunciation control fixes a word with a rule, through a pronunciation dictionary or International Phonetic Alphabet (IPA) tags, so you never retrain to correct a brand name.

No skipped, repeated or hallucinated words

No skipped, repeated or hallucinated words

Autoregressive speech models generate audio one token at a time, and they can loop, drop a word or add one. It’s the most common complaint we hear about open-source local models. For a medication reminder or a robot’s safety instruction, a fluent voice that drops the word “not” is worse than a robotic one that says every word, so we test faithfulness separately from naturalness.

Autoregressive speech models generate audio one token at a time, and they can loop, drop a word or add one. It’s the most common complaint we hear about open-source local models. For a medication reminder or a robot’s safety instruction, a fluent voice that drops the word “not” is worse than a robotic one that says every word, so we test faithfulness separately from naturalness.

Why conversational TTS is a different problem from narration

Why conversational TTS is a different problem from narration

Narration TTS is judged on long passages heard once. Conversational TTS is judged on short turns heard straight after the user stops talking, by a user who may interrupt. A model can win an audiobook comparison and still feel wrong in a conversation, because conversation adds three demands: start fast, sound natural on a two-word reply, and stop cleanly.

Narration TTS is judged on long passages heard once. Conversational TTS is judged on short turns heard straight after the user stops talking, by a user who may interrupt. A model can win an audiobook comparison and still feel wrong in a conversation, because conversation adds three demands: start fast, sound natural on a two-word reply, and stop cleanly.

Short, reactive turns need believable rhythm, not audiobook cadence

Short, reactive turns need believable rhythm, not audiobook cadence

Much of what an assistant says is short: “Sure, one sec.” “Got it, 7:15 tomorrow.” Audiobook cadence makes these lines sound read from a card. Short turns also give the model little text to infer prosody from, so it leans on the reference voice. We recommend cloning from natural, continuous speech, such as a monologue or conversation, so the model picks up conversational rhythm along with timbre.

Much of what an assistant says is short: “Sure, one sec.” “Got it, 7:15 tomorrow.” Audiobook cadence makes these lines sound read from a card. Short turns also give the model little text to infer prosody from, so it leans on the reference voice. We recommend cloning from natural, continuous speech, such as a monologue or conversation, so the model picks up conversational rhythm along with timbre.

Time to first audio is part of sounding human

Time to first audio is part of sounding human

People leave very little silence between speaking turns. Stivers et al. (PNAS, 2009) studied conversations in 10 languages and found speakers in every one keep gaps short, with each language’s average within 250 ms of the cross-language mean. Scientific American’s coverage puts the typical gap around 200 ms and reports that at about 500 ms, listeners began reading a delay as reluctance.

Our working rule: users forgive a slightly less natural voice, but they won’t forgive awkward pauses. Time to first audio belongs on the quality scorecard next to MOS, and every millisecond TTS saves gives the speech recogniser and LLM more room.

People leave very little silence between speaking turns. Stivers et al. (PNAS, 2009) studied conversations in 10 languages and found speakers in every one keep gaps short, with each language’s average within 250 ms of the cross-language mean. Scientific American’s coverage puts the typical gap around 200 ms and reports that at about 500 ms, listeners began reading a delay as reluctance.

Our working rule: users forgive a slightly less natural voice, but they won’t forgive awkward pauses. Time to first audio belongs on the quality scorecard next to MOS, and every millisecond TTS saves gives the speech recogniser and LLM more room.

Interruptions and turn-taking

Interruptions and turn-taking

When a user barges in, the system has to stop and know exactly which word it reached. If the assistant is reading three restaurant options and the user cuts in after the first, an LLM without word-level timestamps records all three as said and may refer to options the user never heard. Precise timestamps let the system trim its context to the last word played. The same timing data drives captions and avatar lip-sync.

When a user barges in, the system has to stop and know exactly which word it reached. If the assistant is reading three restaurant options and the user cuts in after the first, an LLM without word-level timestamps records all three as said and may refer to options the user never heard. Precise timestamps let the system trim its context to the last word played. The same timing data drives captions and avatar lip-sync.

Bar chart: the natural gap between speaking turns is about 200 ms, and about 500 ms starts to feel like hesitation. Below, an on-device voice pipeline next to a cloud pipeline, which adds a network step.
Bar chart: the natural gap between speaking turns is about 200 ms, and about 500 ms starts to feel like hesitation. Below, an on-device voice pipeline next to a cloud pipeline, which adds a network step.

With word timestamps, the assistant’s context stops at the last word the user heard.

The network round trip, the hidden latency on every turn

The network round trip, the hidden latency on every turn

Every cloud TTS call adds a round trip: send text, wait in the provider’s queue, stream audio back. On glasses relaying through a phone over Bluetooth and cellular, that trip can cost more than the model. An African partner told us voice on a foreign server “tends to be plus a second, and up to seven seconds on a bad day”. On-prem or On-device TTS removes the trip and keeps working in tunnels, cold stores and hospital wards.

Every cloud TTS call adds a round trip: send text, wait in the provider’s queue, stream audio back. On glasses relaying through a phone over Bluetooth and cellular, that trip can cost more than the model. An African partner told us voice on a foreign server “tends to be plus a second, and up to seven seconds on a bad day”. On-prem or On-device TTS removes the trip and keeps working in tunnels, cold stores and hospital wards.

Why teams think they must choose between quality and running on the device

Why teams think they must choose between quality and running on the device

Most device teams we speak to have tried at least two of the routes below. Each gives up something the product needs, which is why “quality or control” feels like a law of nature. It holds for models built for data-centre GPUs, and stops holding when the model is designed for the device.

Most device teams we speak to have tried at least two of the routes below. Each gives up something the product needs, which is why “quality or control” feels like a law of nature. It holds for models built for data-centre GPUs, and stops holding when the model is designed for the device.

Cloud TTS: per-minute bills, always connected, and audio leaving the device

Cloud TTS: per-minute bills, always connected, and audio leaving the device

Cloud APIs such as ElevenLabs and Cartesia sound excellent and are easy to start with. The costs arrive later. A per-minute bill on hardware sold once shrinks margin with every active user, so voice gets rate-limited or cut. The product stops talking when the connection drops, and every utterance leaves the device, which triggers privacy reviews.

Cloud APIs such as ElevenLabs and Cartesia sound excellent and are easy to start with. The costs arrive later. A per-minute bill on hardware sold once shrinks margin with every active user, so voice gets rate-limited or cut. The product stops talking when the connection drops, and every utterance leaves the device, which triggers privacy reviews.

Open-source local models: robotic, unreliable, and over the memory and power budget

Open-source local models: robotic, unreliable, and over the memory and power budget

Lightweight engines such as Piper run comfortably on a CPU, but their prosody tends to be flat in dialogue. Larger LLM-based open models sound better, but many need a GPU for real time, exceed a wearable’s RAM or power budget, or hallucinate words under load. Several also ship weights under non-commercial licences. Our guide to open-source speech models covers which ones you can ship.

Lightweight engines such as Piper run comfortably on a CPU, but their prosody tends to be flat in dialogue. Larger LLM-based open models sound better, but many need a GPU for real time, exceed a wearable’s RAM or power budget, or hallucinate words under load. Several also ship weights under non-commercial licences. Our guide to open-source speech models covers which ones you can ship.

What distillation does to prosody

What distillation does to prosody

The usual route to an on-device model is to distill a cloud model into a smaller student and quantise it hard. The architecture was still chosen for a GPU, so the small version spends compute on structures a CPU handles badly. In our experience, the first things to go are prosodic variation, rare pronunciations and stability on unusual inputs. The student learns its teacher’s average delivery, and the average is flat.

The usual route to an on-device model is to distill a cloud model into a smaller student and quantise it hard. The architecture was still chosen for a GPU, so the small version spends compute on structures a CPU handles badly. In our experience, the first things to go are prosodic variation, rare pronunciations and stability on unusual inputs. The student learns its teacher’s average delivery, and the average is flat.

How small models reach human-like quality

How small models reach human-like quality

A device-native model is designed for the target CPU’s compute, memory and battery from the first experiment. It runs on the CPU alone, with no GPU and no network. Its quality comes from architecture choices made for that budget.

A model sounds human when it reads the whole sentence before speaking. That lets it stress the right word and lift its voice at the end of a question. It also has to learn from people talking to each other, because a model trained on audiobook narration will sound like it’s reading aloud. Timing matters too: people leave about 200 ms between turns so a voice that takes longer to start loses the conversation.

The model should be tuned for the one voice and job it has. Small models get there by starting from the device. You fix the chip, threads, memory and battery first, then build the best sounding model that fits, and leave room for speech recognition and the LLM on the same cores.

Shrinking a large cloud model until it fits works in the opposite order, with the device budget considered last. To judge the result, compare models at the same compute cost on word error rate, speaker similarity, prosody heard in context and time to first audio on the target chip.

A device-native model is designed for the target CPU’s compute, memory and battery from the first experiment. It runs on the CPU alone, with no GPU and no network. Its quality comes from architecture choices made for that budget.

A model sounds human when it reads the whole sentence before speaking. That lets it stress the right word and lift its voice at the end of a question. It also has to learn from people talking to each other, because a model trained on audiobook narration will sound like it’s reading aloud. Timing matters too: people leave about 200 ms between turns so a voice that takes longer to start loses the conversation.

The model should be tuned for the one voice and job it has. Small models get there by starting from the device. You fix the chip, threads, memory and battery first, then build the best sounding model that fits, and leave room for speech recognition and the LLM on the same cores.

Shrinking a large cloud model until it fits works in the opposite order, with the device budget considered last. To judge the result, compare models at the same compute cost on word error rate, speaker similarity, prosody heard in context and time to first audio on the target chip.

How to measure conversational TTS quality honestly

How to measure conversational TTS quality honestly

Before you read anyone’s benchmark, including ours, it helps to know how TTS quality numbers are produced and where they mislead. Every common metric captures one aspect of quality and misses others. Used together, with their limits in mind, they give a fair picture. Used alone, each one can be gamed.

Run your own listening tests on your own prompts, treat accuracy metrics as a floor rather than a target, and insist on device numbers that name the hardware.

Before you read anyone’s benchmark, including ours, it helps to know how TTS quality numbers are produced and where they mislead. Every common metric captures one aspect of quality and misses others. Used together, with their limits in mind, they give a fair picture. Used alone, each one can be gamed.

Run your own listening tests on your own prompts, treat accuracy metrics as a floor rather than a target, and insist on device numbers that name the hardware.

Mean Opinion Score (MOS)

Mean Opinion Score (MOS)

Mean opinion score (MOS) asks listeners to rate samples from 1 to 5. Comparative MOS (CMOS) asks them to compare two samples side by side, which is more sensitive to small differences. Both depend heavily on who listens and what they’re told.

Mean opinion score (MOS) asks listeners to rate samples from 1 to 5. Comparative MOS (CMOS) asks them to compare two samples side by side, which is more sensitive to small differences. Both depend heavily on who listens and what they’re told.

WER, and why optimising for it flattens prosody

WER, and why optimising for it flattens prosody

Word error rate (WER) runs generated speech through a speech recogniser and counts the words it gets wrong. It’s a useful check on faithful text generation: it catches skipped, repeated and invented words. It’s a poor guide to naturalness, because a slow, evenly stressed monotone is easy for a recogniser to transcribe. Push a model to minimise WER and it learns to sound careful and flat. Use WER as a pass/fail floor, then judge naturalness by ear.

Word error rate (WER) runs generated speech through a speech recogniser and counts the words it gets wrong. It’s a useful check on faithful text generation: it catches skipped, repeated and invented words. It’s a poor guide to naturalness, because a slow, evenly stressed monotone is easy for a recogniser to transcribe. Push a model to minimise WER and it learns to sound careful and flat. Use WER as a pass/fail floor, then judge naturalness by ear.

Speaker similarity

Speaker similarity

For cloned and custom voices, you also need to know whether the output sounds like the target speaker. Similarity MOS (SMOS) asks listeners. The common objective measure is cosine similarity between speaker embeddings from a verification model such as one built on WavLM. Higher is closer. Check it on conversational prompts as well as read speech, since some models hold the voice on long sentences and drift on short replies.

For cloned and custom voices, you also need to know whether the output sounds like the target speaker. Similarity MOS (SMOS) asks listeners. The common objective measure is cosine similarity between speaker embeddings from a verification model such as one built on WavLM. Higher is closer. Check it on conversational prompts as well as read speech, since some models hold the voice on long sentences and drift on short replies.

Device metrics: RTF, memory, power and time to first audio on a named CPU

Device metrics: RTF, memory, power and time to first audio on a named CPU

These are the hardware requirements that decide whether the model ships:

  • Real-time factor (RTF): generation time divided by audio duration. Below 1.0 means faster than real time.

  • Peak RAM: the most memory the model uses while generating, in MB.

  • Power: average draw during speech, in mW, and what it does to battery life.

  • Time to first audio: milliseconds from text in to first sound out.

Each number is only meaningful with the device, CPU, thread count, quantisation level and test date attached. A figure without them can’t be reproduced.

These are the hardware requirements that decide whether the model ships:

  • Real-time factor (RTF): generation time divided by audio duration. Below 1.0 means faster than real time.

  • Peak RAM: the most memory the model uses while generating, in MB.

  • Power: average draw during speech, in mW, and what it does to battery life.

  • Time to first audio: milliseconds from text in to first sound out.

Each number is only meaningful with the device, CPU, thread count, quantisation level and test date attached. A figure without them can’t be reproduced.

Human-like speech on a CPU benchmark

Human-like speech on a CPU benchmark

How the Neuphonic family of models compare to the market. Below are the figures we have measured, with their conditions.

How the Neuphonic family of models compare to the market. Below are the figures we have measured, with their conditions.

Model
Active parameters
Hardware
Speed
WER (on Seed TTS eval)
Voice Cloning
Languages
NeuTTS-2M60MCPUReal time<0.5%Yes7
NeuTTS Nano (OSS)120MCPUReal time<0.5%Yes4
NeuTTS Max1.7Bn (Total)Accelerator/GPUReal time<0.5%Yes44
Kokoro (OSS)82M (Total)CPUReal time<0.5%No10 (Single language)
Phonon-40M Multilingual (OSS)40MCPUReal time<0.5%No (needs embeddings from online)5

Neuphonic model results against Kokoro and Phonon

Check our numbers: open weights and a reproducible benchmark

Check our numbers: open weights and a reproducible benchmark

NeuTTS Air and Nano weights are on Hugging Face and the code is on GitHub. The basic example clones a voice from a short clip and synthesises speech on your laptop’s CPU in a few commands.

NeuTTS Air and Nano weights are on Hugging Face and the code is on GitHub. The basic example clones a voice from a short clip and synthesises speech on your laptop’s CPU in a few commands.

Custom voices making it yours without losing naturalness

Custom voices making it yours without losing naturalness

“Custom” means two different things in TTS, and teams often ask for one when they need the other. A custom voice changes who the model sounds like. A custom model changes what the model is good at: your vocabulary, your languages, your chip. Many products need both. Both can cost naturalness if done carelessly. If you’re working from our open weights, our guide to fine-tuning open-source TTS covers the hands-on steps.

“Custom” means two different things in TTS, and teams often ask for one when they need the other. A custom voice changes who the model sounds like. A custom model changes what the model is good at: your vocabulary, your languages, your chip. Many products need both. Both can cost naturalness if done carelessly. If you’re working from our open weights, our guide to fine-tuning open-source TTS covers the hands-on steps.

Custom voice cloning from a short reference

Custom voice cloning from a short reference

NeuTTS clones a voice from a short reference recording and its transcript, at inference time. There’s no training run per voice and no per-voice fee, so one model can produce any number of voices: a brand voice for every product line, a cast of characters for a toy, a persona per market.

NeuTTS clones a voice from a short reference recording and its transcript, at inference time. There’s no training run per voice and no per-voice fee, so one model can produce any number of voices: a brand voice for every product line, a cast of characters for a toy, a persona per market.

Keeping prosody intact while you customise

Keeping prosody intact while you customise

Fine-tuning on narrow, carefully read domain data is the fastest way to make a model sound flat. Keep expressive conversational data in the training mix alongside domain data. Hold out a set of conversational prompts, including short replies and questions, and listen to them at every checkpoint. And judge checkpoints by listening tests, using WER only to confirm no words are dropped.

Fine-tuning on narrow, carefully read domain data is the fastest way to make a model sound flat. Keep expressive conversational data in the training mix alongside domain data. Hold out a set of conversational prompts, including short replies and questions, and listen to them at every checkpoint. And judge checkpoints by listening tests, using WER only to confirm no words are dropped.

Owning the voice so it can’t be retired or repriced

Owning the voice so it can’t be retired or repriced

A brand voice is a long-lived asset. Device makers pick a voice partner for design cycles that can run one to five years, and a voice that disappears mid-cycle means rework and a lost identity. With our models you keep the weights and the voices, we own our voice IP outright so nothing can be pulled from under you, and we retain no data.

A brand voice is a long-lived asset. Device makers pick a voice partner for design cycles that can run one to five years, and a voice that disappears mid-cycle means rework and a lost identity. With our models you keep the weights and the voices, we own our voice IP outright so nothing can be pulled from under you, and we retain no data.

Shipping conversational TTS on your device

Shipping conversational TTS on your device

A model that sounds good in a notebook still has to fit the product. Shipping means fitting three budgets at once: latency, memory and power, and the bill of materials. It also means deciding what happens when the device does have a connection and the task needs more than the local model.

The good news is that these budgets reinforce each other. A model designed for the CPU uses less power, needs no new chip and removes the network leg from every turn.

A model that sounds good in a notebook still has to fit the product. Shipping means fitting three budgets at once: latency, memory and power, and the bill of materials. It also means deciding what happens when the device does have a connection and the task needs more than the local model.

The good news is that these budgets reinforce each other. A model designed for the CPU uses less power, needs no new chip and removes the network leg from every turn.

Conclusion

Conclusion

Human-like conversational text to speech comes down to prosody that follows meaning, pauses that come from context, correct names and numbers and no invented words, delivered inside a natural turn gap. Small models can do all of it when they’re designed for the device’s CPU from the first experiment.

Measure before you trust. Compare MOS only within one test, treat WER as a faithfulness floor, name the embedding model behind any similarity score, and ask for RTF, memory and power on a named CPU.

The open NeuTTS models run on your laptop today. Download them from Hugging Face, or book an on-device evaluation and we’ll benchmark NeuTTS on your hardware with your prompts.

Human-like conversational text to speech comes down to prosody that follows meaning, pauses that come from context, correct names and numbers and no invented words, delivered inside a natural turn gap. Small models can do all of it when they’re designed for the device’s CPU from the first experiment.

Measure before you trust. Compare MOS only within one test, treat WER as a faithfulness floor, name the embedding model behind any similarity score, and ask for RTF, memory and power on a named CPU.

The open NeuTTS models run on your laptop today. Download them from Hugging Face, or book an on-device evaluation and we’ll benchmark NeuTTS on your hardware with your prompts.

#conversationalAI

#conversationalAI

#conversationalAI

#texttospeech

#texttospeech

#texttospeech

#voiceAI

#voiceAI

#voiceAI

#ondeviceAI

#ondeviceAI

#ondeviceAI

FAQ

Everything you need to know about human-like conversational text to speech.

Prosody, intonation, natural pausing, correct pronunciation and faithful output. Human-sounding speech stresses the right words, rises on questions, pauses where meaning breaks rather than at every comma, reads numbers and names correctly, and never skips or invents words. In conversation, a fast start matters too, because long silences sound like hesitation.

Yes, if the model is designed for it. Neuphonic’s NeuTTS models run on the CPU alone with no GPU or network, and we’ve shown them working offline on a Raspberry Pi, an NVIDIA Jetson and a Samsung Galaxy S24 in airplane mode. Quality depends on how the model was built, not on whether it’s local.

It can be when the model was designed for the device rather than compressed from a cloud model. Compression tends to flatten prosody. Device-native models target naturalness at their final size from the start. The fair test is a blind listening comparison on your own prompts, run on your production hardware.

Narration reads long passages where latency doesn’t matter. Conversational TTS speaks short, reactive turns, often streamed from a language model, while a user waits. It must sound natural on two-word replies, start speaking within a fraction of a second, and stop cleanly at the exact word when the user interrupts.

Some does. Many open models sound flat in conversation, occasionally skip or invent words, or exceed the memory and power budget of production hardware. Test them on your own conversational prompts and device. Neuphonic publishes open weights for NeuTTS so teams can run exactly that comparison themselves.

A custom voice changes who the model sounds like, usually by cloning from a short reference recording with no training. A custom model changes what it’s good at: domain vocabulary, extra languages or dialects, or optimisation for a specific chip. Custom models involve fine-tuning, often with engineering support on your hardware.

Mostly with listening tests: MOS ratings from 1 to 5, or CMOS side-by-side comparisons. Scores from different studies aren’t comparable because listeners and instructions differ. Word error rate checks accuracy, not naturalness. Device teams should also measure real-time factor, memory, power and time to first audio on a named CPU.

FAQ

Everything you need to know about human-like conversational text to speech.

Prosody, intonation, natural pausing, correct pronunciation and faithful output. Human-sounding speech stresses the right words, rises on questions, pauses where meaning breaks rather than at every comma, reads numbers and names correctly, and never skips or invents words. In conversation, a fast start matters too, because long silences sound like hesitation.

Yes, if the model is designed for it. Neuphonic’s NeuTTS models run on the CPU alone with no GPU or network, and we’ve shown them working offline on a Raspberry Pi, an NVIDIA Jetson and a Samsung Galaxy S24 in airplane mode. Quality depends on how the model was built, not on whether it’s local.

It can be when the model was designed for the device rather than compressed from a cloud model. Compression tends to flatten prosody. Device-native models target naturalness at their final size from the start. The fair test is a blind listening comparison on your own prompts, run on your production hardware.

Narration reads long passages where latency doesn’t matter. Conversational TTS speaks short, reactive turns, often streamed from a language model, while a user waits. It must sound natural on two-word replies, start speaking within a fraction of a second, and stop cleanly at the exact word when the user interrupts.

Some does. Many open models sound flat in conversation, occasionally skip or invent words, or exceed the memory and power budget of production hardware. Test them on your own conversational prompts and device. Neuphonic publishes open weights for NeuTTS so teams can run exactly that comparison themselves.

A custom voice changes who the model sounds like, usually by cloning from a short reference recording with no training. A custom model changes what it’s good at: domain vocabulary, extra languages or dialects, or optimisation for a specific chip. Custom models involve fine-tuning, often with engineering support on your hardware.

Mostly with listening tests: MOS ratings from 1 to 5, or CMOS side-by-side comparisons. Scores from different studies aren’t comparable because listeners and instructions differ. Word error rate checks accuracy, not naturalness. Device teams should also measure real-time factor, memory, power and time to first audio on a named CPU.

FAQ

Everything you need to know about human-like conversational text to speech.

Prosody, intonation, natural pausing, correct pronunciation and faithful output. Human-sounding speech stresses the right words, rises on questions, pauses where meaning breaks rather than at every comma, reads numbers and names correctly, and never skips or invents words. In conversation, a fast start matters too, because long silences sound like hesitation.

Yes, if the model is designed for it. Neuphonic’s NeuTTS models run on the CPU alone with no GPU or network, and we’ve shown them working offline on a Raspberry Pi, an NVIDIA Jetson and a Samsung Galaxy S24 in airplane mode. Quality depends on how the model was built, not on whether it’s local.

It can be when the model was designed for the device rather than compressed from a cloud model. Compression tends to flatten prosody. Device-native models target naturalness at their final size from the start. The fair test is a blind listening comparison on your own prompts, run on your production hardware.

Narration reads long passages where latency doesn’t matter. Conversational TTS speaks short, reactive turns, often streamed from a language model, while a user waits. It must sound natural on two-word replies, start speaking within a fraction of a second, and stop cleanly at the exact word when the user interrupts.

Some does. Many open models sound flat in conversation, occasionally skip or invent words, or exceed the memory and power budget of production hardware. Test them on your own conversational prompts and device. Neuphonic publishes open weights for NeuTTS so teams can run exactly that comparison themselves.

A custom voice changes who the model sounds like, usually by cloning from a short reference recording with no training. A custom model changes what it’s good at: domain vocabulary, extra languages or dialects, or optimisation for a specific chip. Custom models involve fine-tuning, often with engineering support on your hardware.

Mostly with listening tests: MOS ratings from 1 to 5, or CMOS side-by-side comparisons. Scores from different studies aren’t comparable because listeners and instructions differ. Word error rate checks accuracy, not naturalness. Device teams should also measure real-time factor, memory, power and time to first audio on a named CPU.

Bring Neuphonic into your product.

Building something with voice AI? Talk to us about your use case,

deployment requirements, and how Neuphonic can fit into your stack.

Building something with voice AI? Talk to us about your use case,

deployment requirements, and how Neuphonic can fit into your stack.