Speed and price
Time to first audio in milliseconds and price per million characters, in two separate tables: our own measurement, and results read from the Coval leaderboard. Our measurement: Speakvora Phonon voices reached first audio in 44 ms at the median (66 ms at the 95th percentile), $3.00 per million characters. Measured on November 21, 2026 from AWS us-east-1, one stream at a time, on the direct lane path, with any leading silence included.
Where the speed figure comes from: it is our own vendor-run test (November 21, 2026), not an independent benchmark. The rerunnable method and the raw data are in the public tts-latency-benchmark repository, and every price we compare is in the open TTS price index. Speed figures that other vendors publish, for example Deepgram or Cartesia, use different methods and are not directly comparable with ours.
Table 1. Our measurement
Whose measurement: ours, made by the team behind Speakvora, not an independent test. Conditions: November 21, 2026, one AWS c7i.large client in us-east-1, one stream at a time, Speakvora called on its direct lane path (not the public two-step API path), leading silence included in time to first audio, 5 English sentences, two runs pooled. The public API first returns a stream address and token in a separate call, which these figures do not include, so the public two-step path may add time. Polly was called in its own region; Speakvora was reached over the public internet.
| Model | Provider | First audio, median ms | 95th percentile, ms | Arrival + leading silence, median ms | Price per 1M chars | Source |
|---|---|---|---|---|---|---|
| Phonon voices | Speakvora | 44 | 66 | 31 + 12 | $3.00 | Our measurement, Nov 21, 2026 (raw data) |
| Pocket voices | Speakvora | 90 | 141 | 76 + 13 | $3.00 | Our measurement, Nov 21, 2026 (raw data) |
| Polly Standard | Amazon | 43 | 85 | 32 + 9 | $4.00 | Our measurement, Nov 21, 2026 (raw data) |
| Polly Neural | Amazon | 144 | 157 | 36 + 107 | $16.00 | Our measurement, Nov 21, 2026 (raw data) |
| Polly Generative | Amazon | 186 | 212 | 135 + 49 | $30.00 | Our measurement, Nov 21, 2026 (raw data) |
First audio = arrival of the first audio bytes plus the silence before the first audible sound (our implementation of the method Coval describes). The split in column five shows where the time goes: Phonon is about 31 ms of arrival and 12 ms of silence; Polly Neural is 36 ms of arrival and 107 ms of silence, so most of its higher headline number is silence at the start of the audio. Split source: the tts-latency-benchmark README and METHOD, same run. Polly Standard was level with Phonon in this test (43 ms against 44 ms).
Not measured: The 44 ms figure is for catalog Phonon voices. We have not published a time to first audio for new lines in your own cloned voice yet; the first own-voice request after the voice server has been idle can take a minute while it starts.
Table 2. Coval leaderboard results (their method)
Whose measurement: Coval's, from the Coval TTS leaderboard, not ours. These figures were read on October 30, 2026 (their last 24 hours at that time). We tried to read the page again on November 22, 2026 and its table did not load for our fetch, so these are an older reading, not re-read. Their client, location, window and runs differ from Table 1, so the two tables are not the same test and our rows are deliberately not ranked against this list. The price column is the one Coval lists.
| Model | Provider | First audio, median ms | 95th percentile, ms | Price per 1M chars | Source |
|---|---|---|---|---|---|
| vui | Fluxions | 39 | 68 | $9.00 | Coval leaderboard, read Oct 30, 2026 |
| Default | Gradium | 56 | 104 | not listed | Coval leaderboard, read Oct 30, 2026 |
| TTS Flash 2 | Inworld AI | 56 | 111 | $15.00 | Coval leaderboard, read Oct 30, 2026 |
| PhantomZ 4.0 | Deepdub | 60 | 85 | not listed | Coval leaderboard, read Oct 30, 2026 |
| Qwen3 TTS Fast | Nari Labs | 65 | 97 | $10.00 | Coval leaderboard, read Oct 30, 2026 |
| Simba 3.2 | SpeechifyAI | 93 | 127 | $10.00 | Coval leaderboard, read Oct 30, 2026 |
| Simba 3.0 | SpeechifyAI | 99 | 131 | $10.00 | Coval leaderboard, read Oct 30, 2026 |
| Qwen3 TTS 1.7b | Baseten | 105 | 133 | not listed | Coval leaderboard, read Oct 30, 2026 |
| Palabra TTS | Palabra | 118 | 315 | $12.00 | Coval leaderboard, read Oct 30, 2026 |
| Falcon 2 | Murf | 134 | 169 | $10.00 | Coval leaderboard, read Oct 30, 2026 |
| TTS 2 | Inworld AI | 158 | 189 | $25.00 | Coval leaderboard, read Oct 30, 2026 |
| drama3-preview | Fish Audio | 167 | 223 | not listed | Coval leaderboard, read Oct 30, 2026 |
| Eleven v4 Turbo | ElevenLabs | 189 | 218 | $40.00 | Coval leaderboard, read Oct 30, 2026 |
| Assemblyai TTS 1 | AssemblyAI | 204 | 316 | $15.00 | Coval leaderboard, read Oct 30, 2026 |
| Flash v2.5 | ElevenLabs | 206 | 289 | $40.00 | Coval leaderboard, read Oct 30, 2026 |
| Flux TTS | Deepgram | 206 | 361 | not listed | Coval leaderboard, read Oct 30, 2026 |
| Airy TTS | Out of Set | 218 | 487 | $3.67 | Coval leaderboard, read Oct 30, 2026 |
| Lightning v3.1 Pro | Smallest | 263 | 448 | $19.50 | Coval leaderboard, read Oct 30, 2026 |
| TTS Rt v2 | Soniox | 272 | 332 | $14.10 | Coval leaderboard, read Oct 30, 2026 |
| Sonic 3.6 | Cartesia | 272 | 333 | not listed | Coval leaderboard, read Oct 30, 2026 |
| TTS RT v1 | Soniox | 280 | 331 | $14.10 | Coval leaderboard, read Oct 30, 2026 |
| Sonic 3.5 | Cartesia | 280 | 424 | not listed | Coval leaderboard, read Oct 30, 2026 |
| Aura 2 | Deepgram | 286 | 524 | $30.00 | Coval leaderboard, read Oct 30, 2026 |
| Grok TTS | SpaceXAI | 324 | 518 | $15.00 | Coval leaderboard, read Oct 30, 2026 |
| Daytona TTS | Guava | 334 | 441 | not listed | Coval leaderboard, read Oct 30, 2026 |
| S2.1 Pro | Fish Audio | 362 | 493 | $15.00 | Coval leaderboard, read Oct 30, 2026 |
| Mist v3 | Rime | 367 | 484 | $30.00 | Coval leaderboard, read Oct 30, 2026 |
| Coda | Rime | 381 | 543 | $50.00 | Coval leaderboard, read Oct 30, 2026 |
| Chirp 3 HD | 547 | 867 | $30.00 | Coval leaderboard, read Oct 30, 2026 | |
| Aura 2 En | Cloudflare | 547 | 1023 | not listed | Coval leaderboard, read Oct 30, 2026 |
| Qwen3 TTS Flash Realtime | Alibaba | 723 | 778 | $13.00 | Coval leaderboard, read Oct 30, 2026 |
| S2.1 Pro Free | Fish Audio | 758 | 939 | not listed | Coval leaderboard, read Oct 30, 2026 |
How we measured
- Client: one AWS c7i.large in us-east-1, over the public internet.
- Speakvora: Phonon (Delia, Easton) and Pocket (Jasper, Marisol) voices streamed from our Mac mini through its public Tailscale Funnel address. The streaming endpoint was called directly with a pre-issued token, so the call that first returns a stream address is not in these figures.
- Amazon Polly: Standard (Joanna), Neural (Joanna) and Generative (Ruth), called in us-east-1 with the AWS SDK and PCM output.
- One stream at a time: a single request on a kept-alive connection, never several at once. This is not a load test and says nothing about how many streams a server can carry.
- 5 English sentences, 12 requests per sentence per service in each of 2 runs (120 requests per voice), services interleaved in shuffled order, after warm-up requests. 240 requests per Speakvora voice family (two voices), 120 per Polly engine.
- Time to first audio is the arrival of the first audio chunk plus the silence before the first audible sample (10 ms window, RMS above 0.01). It is our own implementation of the method Coval describes, so we call it Coval's method as we read it, not Coval's result.
What this does not show
- Concurrency: every figure is one stream at a time. It says nothing about speed when many callers share a server.
- Distance: your own network adds its own delay; our client sat in the same cloud region as Amazon Polly and reached our Mac mini over the internet.
- Other voices, long text, languages other than English, your own cloned voice and audio quality. Listen to the samples and test with your own text.
- Our own test, not a third-party benchmark. There is no latency or uptime guarantee.
Raw data
Every timed request of both runs, with arrival time and leading silence for each: run 1 (JSON) and run 2 (JSON). The pooled medians and 95th percentiles in the table above are computed from these two files. For a one-on-one test with audio to compare by ear, see Speakvora against Amazon Polly Neural.
Part of our text to speech API guide: voices, price, speed and code samples.
Quick answers
How much is free?
Every account gets 100,000 characters of speech free every month, with no credit card, plus 60 minutes of speech to text every month. Both renew on the 1st of every month (UTC).
How fast is Speakvora text to speech?
Measured on our own Mac mini on 2026-11-21: first audio in about 44 ms with Phonon voices and 90 ms with Pocket voices (measured 2026-11-21 with the client in AWS us-east-1; your own network adds its own delay).
What does it cost after the free allowance?
Text to speech is $3 per million characters and speech to text is $0.0015 per minute. Industry agents are a one-time payment from $214 with no per-call, per-text or per-minute fee from Speakvora.
Are the voices human?
No. They are AI voices chosen by ear to sound natural; you can listen to each one before you buy.