On a single NVIDIA V100 GPU, the HiFi-GAN V3 vocoder generated audio 1,186.80 times faster than real time, while an autoregressive WaveNet managed 0.003 times real time on the same card (Kong et al., HiFi-GAN 2020). That gap, more than five orders of magnitude, is why speech synthesis went from a data-center luxury to something a phone can do. The bottleneck has since moved: the original F5-TTS paper, a large zero-shot model, reported 0.15 seconds of GPU time per second of speech (Chen et al., F5-TTS 2024), and a streaming test on one A100 measured first audio anywhere from 102 ms to 1,643 ms depending on the system (KTH, VoXtream 2025). On the price side, major cloud list prices run from $4 to $160 per million characters, a 40-fold spread for the same unit of text. We aggregated data from peer-reviewed and arXiv papers by KTH, Kyutai and Apple, NVIDIA’s Riva performance documentation, and the official pricing pages of Google Cloud, Amazon Web Services, Microsoft Azure and OpenAI; quality scores are covered separately in our synthetic speech MOS statistics.
A note on the unit that runs through this article: the real-time factor (RTF) normally means compute time divided by audio duration, so an RTF of 0.1 means one second of speech takes 0.1 seconds to generate, and lower is better. Several sources invert the ratio and report how many seconds of audio come out per second of compute (NVIDIA calls this RTFX, and some papers write it as “times faster than real time”), where higher is better. In the tables, RTF means the lower-is-faster convention unless a row says “times real time”, “RTFX” or “higher is faster”.
TL;DR
- FastSpeech 2 cut RTF from 0.932 for Transformer TTS to 0.0195, a 47.8x speedup (Ren et al., FastSpeech 2 2020).
- Matcha-TTS with 2 sampling steps measured an RTF of 0.015 including the vocoder (Mehta et al., Matcha-TTS 2023).
- F5-TTS reported RTF 0.15 at 16 steps versus 0.92 for CosyVoice in its comparison table (Chen et al., F5-TTS 2024).
- VoXtream reached 102 ms first-packet latency on an A100, while CosyVoice2 in full-stream mode took 1,643 ms in the same test (KTH, VoXtream 2025).
- Kyutai’s DSM-TTS latency rose from 150 ms at batch size 1 to 403 ms at batch size 64, while throughput reached 137.3 times real time (Kyutai, DSM-TTS 2025).
- NVIDIA Riva’s FastPitch plus HiFi-GAN pipeline went from 22 ms first-audio latency at 1 stream to 323 ms at 32 streams (NVIDIA Riva 2.16.0 documentation, vendor-run).
- Apple’s on-device system ran 1.47 to 3.03 times faster than real time on two mobile devices (Apple, On-device neural speech synthesis 2021).
- NICT’s Mobile PresenTra reached a batch RTF of about 0.3 on a middle-range Google Pixel 5 (Okamoto et al., Mobile PresenTra, Interspeech 2024).
- Cutting SupertonicTTS from 8 to 4 sampling steps raised word error rate from 2.818% to 11.43% (Supertone, SupertonicTTS 2025).
- Fast F5-TTS ran about 4 times faster than the 32-step baseline while word error rate moved from 2.37% to 2.45% (Zheng et al., Interspeech 2025).
- Google Cloud lists $4 per 1M characters for Standard voices and $160 for Studio voices (Google Cloud Text-to-Speech pricing 2026).
- Amazon equates 1M characters with about 23 hours 8 minutes of speech, which puts Polly Generative at about $1.30 per hour of audio (AWS Amazon Polly pricing 2026).
1. Real-Time Factor by Model Family on GPU
The first big speed jump came from dropping autoregression in the acoustic model, the part that turns text into a mel spectrogram (a time-frequency picture of speech). FastSpeech 2 measured an RTF of 0.0195 against 0.932 for Transformer TTS on the same V100 server, a 47.8x inference speedup (Ren et al., FastSpeech 2). Once a model predicts all frames in parallel, it no longer has to wait on its own previous output, frame by frame.
Newer families did not undo that gain, but they did spread it out. Diffusion and flow-matching models such as Grad-TTS and Matcha-TTS trade a few extra network passes for better quality, and in the Matcha-TTS benchmark on RTX 3090 GPUs, all of the tested systems stayed under an RTF of 0.08 including vocoding (Mehta et al., Matcha-TTS). The interesting number there is not the mean but the spread: VITS posted 0.074 with a standard deviation of 0.083, larger than the mean itself, which signals very uneven speed from sentence to sentence.
| Metric | Value | Source |
|---|---|---|
| Transformer TTS RTF, V100 GPU, batch size 1 (lower is faster) | 0.932 | Ren et al., FastSpeech 2 2020 |
| FastSpeech 2 RTF and speedup over Transformer TTS, same setup | 0.0195 (47.8x) | Ren et al., FastSpeech 2 2020 |
| FastSpeech 2s (text straight to waveform) RTF and speedup | 0.0180 (51.8x) | Ren et al., FastSpeech 2 2020 |
| FastSpeech 2 RTF in the Matcha-TTS test, including vocoder, RTX 3090 | 0.010 (standard deviation 0.004) | Mehta et al., Matcha-TTS 2023 |
| Matcha-TTS RTF with 2 sampling steps vs 10 sampling steps | 0.015 vs 0.038 | Mehta et al., Matcha-TTS 2023 |
| Grad-TTS RTF with 10 sampling steps | 0.049 | Mehta et al., Matcha-TTS 2023 |
| VITS RTF in the Matcha-TTS test (mean and standard deviation) | 0.074 (0.083) | Mehta et al., Matcha-TTS 2023 |
| StyleTTS 2 RTF vs VITS RTF in the StyleTTS 2 test | 0.0185 vs 0.0599 | Li et al., StyleTTS 2 2023 |
Most recent available data: FastSpeech 2 (2020), StyleTTS 2 (2023) and Matcha-TTS (2023), all older than three years. RTF values from different papers are not directly comparable, because GPUs, batch sizes and whether the vocoder is included all differ. The Matcha-TTS authors also report that their 2-step model becomes competitive with FastSpeech 2 on longer utterances.
2. Vocoders: Where the Compute Goes
A vocoder converts the mel spectrogram into an audible waveform, and for years it was the expensive half of the pipeline. WaveNet produced 0.07 kHz of audio on a V100 GPU, which is 0.003 times real time, while HiFi-GAN V1 produced 3,701 kHz, or 167.86 times real time (Kong et al., HiFi-GAN). The point is not only that GAN vocoders are fast; it is that the vocoder stopped being the reason a TTS product needed a GPU at all.
The same pattern now repeats inside streaming systems. Apple’s SpeakStream paper shows the acoustic model taking about 28 ms while a Parallel WaveGAN (PWG) vocoder added 152 ms; swapping in a streaming vocoder cut that stage to 13 ms (Apple, SpeakStream). In other words, the choice of vocoder alone decided whether first audio arrived in about 40 ms or about 180 ms.
| Metric | Value | Source |
|---|---|---|
| WaveNet (MoL) vs WaveGlow vs MelGAN speed on a V100 GPU, times real time (higher is faster) | 0.003x vs 22.75x vs 645.73x | Kong et al., HiFi-GAN 2020 |
| HiFi-GAN V1 vs HiFi-GAN V3 speed on a V100 GPU, times real time | 167.86x vs 1,186.80x | Kong et al., HiFi-GAN 2020 |
| BigVGAN vocoder RTF vs SupertonicTTS speech autoencoder RTF, RTX 4090 (lower is faster) | 0.0124 vs 0.0006 | Supertone, SupertonicTTS 2025 |
| SpeakStream on an M4 Pro Mac mini, with PWG vs with the streaming VocStream vocoder: vocoder time and total first-audio time | 152 ms vs 13 ms; 180 ms vs 40 ms | Apple, SpeakStream 2025 |
Most recent available data: Kong et al., HiFi-GAN 2020, measured in 32-bit floating point with no further optimization. In the Matcha-TTS test the vocoder alone ran at an RTF of 0.001, so in that setup it was a small fraction of total synthesis time (Matcha-TTS).
3. Large Zero-Shot Models: The Speed Cost of Voice Cloning at Scale
Zero-shot models, which imitate a new voice from a short prompt without retraining, brought speed back into the conversation. The F5-TTS comparison table lists CosyVoice at an RTF of 0.92 and FireRedTTS at 0.84, meaning each second of speech took most of a second to compute (Chen et al., F5-TTS). F5-TTS itself reported 0.15 at 16 sampling steps on an RTX 3090, or about 6.7 times faster than real time (1 divided by 0.15). That is fast enough for batch audiobook work but tight for serving many live users from one card.
Smaller models are closing the gap. SupertonicTTS, with 44M parameters, reported 0.02 on an RTX 4090 and 0.05 on an RTX 3090, against 0.31 for F5-TTS at 32 steps on the RTX 3090 (Supertone, SupertonicTTS). For teams evaluating an AI voice generator, parameter count alone is a poor guide to serving cost: the table below lists DiTTo-TTS at 740M parameters as faster than Voicebox at 330M, though those two figures may not come from the same hardware.
| Metric | Value | Source |
|---|---|---|
| F5-TTS RTF, 16 vs 32 sampling steps, RTX 3090, 10 s of speech (lower is faster) | 0.15 vs 0.31 | Chen et al., F5-TTS 2024 |
| CosyVoice (about 300M parameters) vs FireRedTTS (about 580M) RTF | 0.92 vs 0.84 | Chen et al., F5-TTS 2024 |
| Voicebox (330M) vs VALL-E 2 RTF | 0.64 vs 0.732 | Chen et al., F5-TTS 2024 |
| DiTTo-TTS (740M) vs NaturalSpeech 3 (500M) RTF | 0.162 vs 0.296 | Chen et al., F5-TTS 2024 |
| SupertonicTTS (44M) RTF on RTX 4090 vs RTX 3090 | 0.02 vs 0.05 | Supertone, SupertonicTTS 2025 |
| Real-time multiple at batch size 1 (higher is faster): DSM-TTS vs Chatterbox vs CSM | 3.2 vs 1.8 vs 1.0 | Kyutai, DSM-TTS 2025 |
| Real-time multiple at batch size 1: Dia and Orpheus | 0.7 each (slower than real time) | Kyutai, DSM-TTS 2025 |
Context note: the F5-TTS paper does not say how the baseline RTF values were measured, and they may not share its RTX 3090, so treat rows 2 to 4 as indicative. A later paper measured F5-TTS itself at 0.065 and 0.123 on an RTX 3090 (see section 6). Kyutai’s table uses the inverted convention, where a value below 1.0 means the model cannot keep up with playback.
4. Streaming Time to First Audio and Serving Concurrency
For live use, the number that matters is not RTF but how long the listener waits for the first sound. In one A100 benchmark, VoXtream with torch.compile started speaking after 102 ms, while CosyVoice2 in full-stream mode took 1,643 ms (KTH, VoXtream). Full-stream means the model starts while text is still arriving; a system can have a decent RTF and still a poor first-packet latency if it waits for whole phrases.
Concurrency is the other half. Batching more users onto one GPU raises total throughput but stretches each user’s wait. Kyutai’s DSM-TTS went from 150 ms at batch size 1 to 403 ms at batch size 64, while throughput rose from 3.2 to 137.3 times real time (Kyutai, DSM-TTS). NVIDIA’s own Riva tables show the same curve on a FastPitch plus HiFi-GAN pipeline: first-audio latency grew about 15-fold from 1 to 32 streams while throughput grew about 2.9-fold (NVIDIA Riva TTS performance). How these figures fit a full voice-agent latency budget is covered in our voice agent time-to-first-audio statistics.
| Metric | Value | Source |
|---|---|---|
| VoXtream first-packet latency and RTF with torch.compile, A100, FP16 | 102 ms (RTF 0.17) | KTH, VoXtream 2025 |
| VoXtream first-packet latency and RTF without compilation | 171 ms (RTF 1.00) | KTH, VoXtream 2025 |
| XTTS-v2 first-packet latency, output streaming, without vs with DeepSpeed | 295 ms vs 196 ms | KTH, VoXtream 2025 |
| CosyVoice2 first-packet latency and RTF, full-stream mode, same A100 test | 1,643 ms (RTF 0.85) | KTH, VoXtream 2025 |
| DSM-TTS latency at batch size 1 vs 32 vs 64 | 150 ms vs 380 ms vs 403 ms | Kyutai, DSM-TTS 2025 |
| DSM-TTS throughput at batch size 64, times real time | 137.3x | Kyutai, DSM-TTS 2025 |
| Riva FastPitch plus HiFi-GAN average first-audio latency, 1 stream vs 32 streams | 22 ms vs 323 ms | NVIDIA Riva 2.16.0 documentation (vendor-run) |
| Riva throughput (RTFX, higher is faster), 1 stream vs 32 streams | 150.8 vs 440 | NVIDIA Riva 2.16.0 documentation (vendor-run) |
Context note: the CosyVoice project README claims latency “as low as 150ms” for its newer Fun-CosyVoice 3.0 model (FunAudioLLM CosyVoice on GitHub), a project claim for a different version than the one KTH measured. We do not name a GPU for the Riva rows: they are the first results table on NVIDIA’s page, whose GPU list starts with A100, but the page text we could read did not tie each table to its GPU label.
5. CPU, Mobile and Edge Hardware
Off the GPU, speed margins shrink fast. HiFi-GAN V1, which ran 167.86 times faster than real time on a V100, managed only 1.43 times real time on a laptop Intel i7 CPU, while WaveGlow fell to 0.21 times, too slow for live playback (Kong et al., HiFi-GAN). The model that wins on a server is often the wrong one for a laptop, and the smaller HiFi-GAN V3 (13.44 times real time on the same CPU) shows how much architecture choices matter here.
Purpose-built small models push much further. EfficientSpeech, with 266k parameters and 90 MFLOPS, generated mel spectrograms 104.3 times faster than real time on a Raspberry Pi 4, about 1% of the size and computation of compact models such as Mixer-TTS (EfficientSpeech). Apple’s on-device work showed the full pipeline running on two mobile devices, with the first audio buffer playing 158 to 179 ms after the request (Apple, On-device neural speech synthesis). Hardware trends behind this shift are tracked in our edge AI voice processor statistics.
The newest measurements show how far the design choice matters on the same chip. On one thread of an AMD EPYC 7413, a streaming model from Samsung Research reached an RTF of 0.568 and produced its first acoustic token in 0.061 s, against 7.665 and 41.41 s for VALL-E X (Lee et al., Interspeech 2025). NICT reported a batch RTF of about 0.3 and incremental latency of less than 0.5 s on a Google Pixel 5, which it describes as a middle-range phone (Okamoto et al., Mobile PresenTra).
| Metric | Value | Source |
|---|---|---|
| HiFi-GAN V1 vs HiFi-GAN V3 vs WaveGlow speed on a laptop Intel i7 CPU, times real time | 1.43x vs 13.44x vs 0.21x | Kong et al., HiFi-GAN 2020 |
| Apple end-to-end synthesis speed, mobile device A vs device B, times real time | 1.47x vs 3.03x | Apple, On-device neural speech synthesis 2021 |
| Apple customer perceived latency (request to first audio buffer played), device A vs device B | 179 ms vs 158 ms | Apple, On-device neural speech synthesis 2021 |
| Mobile PresenTra on a Google Pixel 5: batch RTF; incremental latency | about 0.3; less than 0.5 s | Okamoto et al. (NICT), Mobile PresenTra, Interspeech 2024 |
| RTF on one AMD EPYC 7413 CPU thread, token generation only (lower is faster): SMAM+INR vs Small-E vs VALL-E X | 0.568 vs 2.404 vs 7.665 | Lee et al. (Samsung Research), Interspeech 2025 |
| Time to first acoustic token on the same CPU thread: SMAM+INR vs Small-E vs VALL-E X | 0.061 s vs 14.34 s vs 41.41 s | Lee et al. (Samsung Research), Interspeech 2025 |
| EfficientSpeech parameters and mel generation speed on a Raspberry Pi 4 | 266k parameters, 104.3x real time | EfficientSpeech 2023 |
| Nix-TTS vs its VITS teacher model, RTF on a Raspberry Pi 3B (above 1.0 is slower than real time) | 1.974 vs 16.50 | Nix-TTS 2022 |
Most recent available data: Kong et al. 2020, Apple 2021, Nix-TTS 2022 and EfficientSpeech 2023 are older than three years; we found no newer peer-reviewed figures for those exact comparisons. Apple’s abstract rounds the mobile figure to “3x faster than real time”, but its table shows device A at 1.47x. The EfficientSpeech figure covers mel generation only, not the vocoder, and the Samsung rows exclude codec processing, so neither is a full speech pipeline.
6. The Speed-Quality Trade-Off
Diffusion and flow-matching models expose speed as a dial: the number of sampling steps (network function evaluations, or NFE). SupertonicTTS kept its word error rate between 2.639% and 2.818% from 8 to 128 steps, but at 4 steps the error rate jumped to 11.43% and the NISQA quality score fell to 2.623 (Supertone, SupertonicTTS). Intelligibility does not fade gradually; it holds steady and then breaks at a cliff, so the cheapest safe setting sits just above it.
That cliff can move without new hardware. On an RTX 3090, F5-TTS at 7 evenly spaced steps and Fast F5-TTS at 7 pruned steps both ran at an RTF of 0.030, but only the choice of where the steps fall took word error rate from 4.16% down to 2.45%, close to the 2.37% of the 32-step baseline at 0.123 (Zheng et al., Interspeech 2025).
The same shape appears in vocoders and streaming. HiFi-GAN V3 ran about 7.07 times faster than V1 on a V100 GPU (26,169 kHz divided by 3,701 kHz), for a MOS drop from 4.36 to 4.05 (Kong et al., HiFi-GAN). CosyVoice 2 called its streaming mode “nearly lossless in typical test cases”, yet on its hard test set word error rate rose from 6.83% offline to 8.08% streaming (CosyVoice 2).
| Metric | Value | Source |
|---|---|---|
| SupertonicTTS at 4 vs 8 sampling steps: RTF; word error rate; NISQA score | 0.006 vs 0.011; 11.43% vs 2.818%; 2.623 vs 3.916 | Supertone, SupertonicTTS 2025 |
| SupertonicTTS at 32 steps (chosen setting) vs 128 steps: RTF; NISQA score (recordings score 4.070) | 0.037 vs 0.140; 4.033 vs 4.070 | Supertone, SupertonicTTS 2025 |
| F5-TTS RTF on an RTX 3090 at 32 vs 16 vs 7 sampling steps | 0.123 vs 0.065 vs 0.030 | Zheng et al., Interspeech 2025 |
| Word error rate on LibriSpeech-PC at 7 steps, evenly spaced F5-TTS vs Fast F5-TTS (32-step baseline 2.37%) | 4.16% vs 2.45% | Zheng et al., Interspeech 2025 |
| Matcha-TTS MOS with 2 vs 10 sampling steps | 3.65 vs 3.84 | Mehta et al., Matcha-TTS 2023 |
| CosyVoice 2 word error rate on the hard test set, offline vs streaming | 6.83% vs 8.08% | CosyVoice 2 2024 |
| StyleTTS 2 RTF with 4 vs 128 diffusion steps | 0.0179 vs 0.0963 | Li et al., StyleTTS 2 2023 |
Context note: NISQA is an automatic, non-intrusive speech quality predictor that estimates a listener-style score without a reference recording; it is not a human listening test. The SupertonicTTS step table does not state its hardware, and its 0.037 at 32 steps differs from the 0.02 reported elsewhere in the same paper.
7. Price per Million Characters and per Hour of Audio
List prices show a market split into tiers rather than a single going rate. Google Cloud charges $4 per 1 million characters for Standard and WaveNet voices but $160 for Studio voices, a 40-fold difference for the same text (Google Cloud Text-to-Speech pricing). Amazon Polly runs from $4 for Standard to $100 for Long-Form voices, and its own example equates 1 million characters with about 23 hours and 8 minutes of speech (AWS Amazon Polly pricing).
Using that AWS conversion (23 hours 8 minutes is about 23.13 hours), Polly Generative at $30 works out to about $1.30 per hour of audio (30 divided by 23.13), and Standard to about $0.17. Token-priced models need a different calculation: Google states that Gemini TTS audio is 25 tokens per second, so one hour is 90,000 audio tokens, and Gemini 2.5 Pro TTS at $20 per million audio tokens comes to $1.80 per hour of output before text input charges. Open-weight models compress this further: Artificial Analysis lists Kokoro 82M on Replicate at $0.65 per million characters (Artificial Analysis text-to-speech leaderboard). For the wider GPU economics behind these prices, see our AI inference cost statistics.
| Metric | Value | Source |
|---|---|---|
| Google Cloud list price per 1M characters: Standard and WaveNet vs Neural2 | $4 vs $16 | Google Cloud Text-to-Speech pricing 2026 |
| Google Cloud list price per 1M characters: Chirp 3 HD vs Instant custom voice vs Studio | $30 vs $60 vs $160 | Google Cloud Text-to-Speech pricing 2026 |
| Amazon Polly list price per 1M characters: Standard, Neural, Generative, Long-Form | $4, $16, $30, $100 | AWS Amazon Polly pricing 2026 |
| Microsoft Azure neural voices, pay as you go vs overage price in the 2,000M-character commitment tier, per 1M characters | $15 vs $7.50 | Microsoft Azure Speech pricing 2026 |
| OpenAI tts-1 (described as optimized for speed) vs tts-1-hd, per 1M characters | $15 vs $30 | OpenAI API pricing 2026 |
| Gemini 2.5 Pro TTS audio output price and derived cost per hour of audio (25 tokens per second) | $20 per 1M audio tokens, about $1.80 per hour | Google Cloud Text-to-Speech pricing 2026 (derived) |
| Amazon Polly Generative and Standard, derived cost per hour of audio | about $1.30 and about $0.17 | AWS Amazon Polly pricing 2026 (derived) |
| Kokoro 82M hosted API price per 1M characters (Replicate) | $0.65 | Artificial Analysis TTS leaderboard 2026 |
Context note: these are list prices as of October 31, 2026, before negotiated discounts. Azure’s commitment tiers start at $960 for 80M characters, which is $12 per million (960 divided by 80). Free tiers differ widely: Polly Standard includes 5 million characters a month, while Polly Generative includes 100 thousand characters a month for the first 12 months. Kokoro’s own model card says the model was trained for about $1,000 using 1,000 A100 80GB GPU hours (Kokoro-82M model card), which shows how little training compute a low-cost voice now needs.
Summary: Text-to-Speech Real-Time Factor and Inference Cost by the Numbers
| Metric | Value | Source |
|---|---|---|
| HiFi-GAN V3 vs WaveNet speed on a V100 GPU, times real time | 1,186.80x vs 0.003x | Kong et al., HiFi-GAN 2020 |
| FastSpeech 2 vs Transformer TTS RTF | 0.0195 vs 0.932 | Ren et al., FastSpeech 2 2020 |
| Matcha-TTS RTF with 2 sampling steps, including vocoder | 0.015 | Mehta et al., Matcha-TTS 2023 |
| F5-TTS RTF at 16 steps vs CosyVoice RTF | 0.15 vs 0.92 | Chen et al., F5-TTS 2024 |
| VoXtream vs CosyVoice2 first-packet latency, same A100 test | 102 ms vs 1,643 ms | KTH, VoXtream 2025 |
| DSM-TTS latency, batch size 1 vs 64 | 150 ms vs 403 ms | Kyutai, DSM-TTS 2025 |
| Riva first-audio latency, 1 vs 32 streams | 22 ms vs 323 ms | NVIDIA Riva 2.16.0 documentation (vendor-run) |
| SpeakStream first audio with PWG vs streaming vocoder | 180 ms vs 40 ms | Apple, SpeakStream 2025 |
| Apple on-device synthesis speed, two mobile devices | 1.47x and 3.03x real time | Apple, On-device neural speech synthesis 2021 |
| EfficientSpeech mel generation on a Raspberry Pi 4 | 104.3x real time | EfficientSpeech 2023 |
| Mobile PresenTra batch RTF on a Google Pixel 5 | about 0.3 | Okamoto et al. (NICT), Mobile PresenTra, Interspeech 2024 |
| SupertonicTTS word error rate at 4 vs 8 sampling steps | 11.43% vs 2.818% | Supertone, SupertonicTTS 2025 |
| F5-TTS word error rate at 7 steps, evenly spaced vs pruned schedule | 4.16% vs 2.45% | Zheng et al., Interspeech 2025 |
| CosyVoice 2 hard-set word error rate, offline vs streaming | 6.83% vs 8.08% | CosyVoice 2 2024 |
| Google Cloud Standard vs Studio voices, per 1M characters | $4 vs $160 | Google Cloud Text-to-Speech pricing 2026 |
| Amazon Polly Long-Form voices, per 1M characters | $100 | AWS Amazon Polly pricing 2026 |
| Speech produced by 1M characters, Amazon’s pricing example | about 23 hours 8 minutes | AWS Amazon Polly pricing 2026 |
| Azure neural voices, pay as you go per 1M characters | $15 | Microsoft Azure Speech pricing 2026 |
| Kokoro 82M hosted price per 1M characters | $0.65 | Artificial Analysis TTS leaderboard 2026 |
Methodology and Sources
- Kong, Kim and Bae: HiFi-GAN (2020), Table 1 speed, MOS and parameter counts on a V100 GPU and a laptop CPU.
- Ren et al.: FastSpeech 2 (2020), Table 2 inference RTF on a V100 server at batch size 1.
- Mehta et al.: Matcha-TTS (2023), RTF including vocoding on RTX 3090 GPUs.
- Li et al.: StyleTTS 2 (2023), RTF overall and by diffusion step count.
- Chen et al.: F5-TTS (2024), RTF on an RTX 3090 for 10 s of speech, with a baseline comparison table.
- Supertone: SupertonicTTS (2025), RTF on RTX 4090 and RTX 3090, and the sampling-step ablation.
- KTH Royal Institute of Technology: VoXtream (2025), first-packet latency and RTF in FP16 on an A100.
- Kyutai: Streaming Sequence-to-Sequence Learning with Delayed Streams Modeling (2025), DSM-TTS latency, RTF and throughput by batch size.
- Apple: SpeakStream (2025), latency on an M4 Pro Mac mini with MLX; On-device neural speech synthesis (2021), mobile measurements.
- FunAudioLLM: CosyVoice 2 (2024), streaming vs offline quality; CosyVoice GitHub README (project claim, 2026).
- EfficientSpeech (2023) and Nix-TTS (2022), small-footprint models on CPU and Raspberry Pi.
- Interspeech papers: Okamoto, Ohtani and Kawai (NICT), Mobile PresenTra (2024), smartphone RTF and latency; Lee et al. (Samsung Research), Efficient Streaming TTS Acoustic Model with Depthwise RVQ Decoding Strategies in a Mamba Framework (2025), single-thread CPU RTF and latency; Zheng et al., Accelerating Flow-Matching-Based Text-to-Speech via Empirically Pruned Step Sampling (2025), F5-TTS speed and error rates by step count.
- NVIDIA: Riva 2.16.0 TTS performance, vendor-run latency and RTFX tables.
- Pricing pages, read on October 31, 2026: Google Cloud Text-to-Speech, Amazon Polly, Microsoft Azure Speech, OpenAI API and the OpenAI tts-1 model page.
- Artificial Analysis text-to-speech leaderboard (2026) and the Kokoro-82M model card (Hugging Face), for open-weight model pricing and training cost.
- Data watch: RTF numbers are not comparable across papers unless hardware, batch size, precision and vocoder inclusion match; even within one paper, SupertonicTTS reports 0.037 and 0.02 for what looks like the same setting. The F5-TTS paper does not state how its baseline RTFs were measured, and F5-TTS speed itself conflicts between sources on the same GPU model: the original paper gives 0.15 at 16 steps and 0.31 at 32 steps on an RTX 3090, while Zheng et al. give 0.065 and 0.123 on an RTX 3090, with no explanation of the gap in the pages we read. Kyutai’s Table 6 does not name its GPU, and the paper prints about 800 ms latency for F5-TTS at 16 steps and about 400 ms at 32 steps, the opposite of what more steps would normally do, so we did not use those two values. The VoXtream measurement and the CosyVoice README figure cover different CosyVoice versions, so they are not directly comparable. The two conversions from characters to audio conflict: Amazon’s example implies about 720 characters per minute (1,000,000 characters divided by 1,388 minutes), while the Kokoro card uses about 1,000, so a per-hour cost based on Kokoro’s ratio would be about 39% higher than one based on Amazon’s (1,000 divided by 720 is about 1.39). Google Cloud labels its Gemini 3.8 TTS models as Preview, with prices scheduled to change on January 1, 2027. The MLPerf Inference rules from MLCommons list Whisper for speech-to-text but no text-to-speech workload, so TTS speed has no neutral industry benchmark there; several of the hardware papers here are more than three years old and are flagged as most recent available data.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when Google Cloud’s January 1, 2027 Gemini TTS price change takes effect.