Disclosure: This article was written with AI assistance and reviewed by EmpirioLabs AI.
Short answer: for the same 174-character script, Qwen Audio 3.1 TTS cost the least, under a tenth of a cent, StepAudio 3 TTS returned its file fastest, and Gemini 3.8 Flash TTS cost the most, under two cents. Qwen Audio 3.0 TTS cost under half a cent. All four run on EmpirioLabs through the same speech endpoint. Listen to each clip below; we are not declaring a winner.
Listen to the four voices
The script: "I opened the letter expecting another polite rejection. Instead, the first line said: congratulations, the job is yours. I read it three times before I let myself believe it." The direction: "Start quiet and a little tired, then grow surprised and joyful by the last sentence."
Results at a glance
| Qwen Audio 3.1 TTS | Qwen Audio 3.0 TTS | Gemini 3.8 Flash TTS | StepAudio 3 TTS | |
|---|---|---|---|---|
| Maker | Alibaba | Alibaba | StepFun | |
| Billing unit | Tokens: text in, 12.5 tokens per second of audio out | Characters | Tokens | Characters |
| Price for this request | $0.0008 | $0.0035 | $0.0180 | $0.0026 |
| Audio length | 16.8 seconds | 13.4 seconds | 15.4 seconds | 12.3 seconds |
| Time to the finished file | 13.4 seconds | 16.6 seconds | 7.5 seconds | 4.4 seconds |
| Voices | 19 | Over 1,000 | 30 in the voice list, over 2,000 in the library | 12 official voices |
| Direction field | instruction | instruction | style_prompt | instruction |
| Streaming | Yes | Yes | Yes | Yes |
How we ran it
Each model read the identical script with the identical direction, passed in the model's own direction field. Each used an English female voice from its own library, returned MP3, and kept every other setting at its default; Qwen Audio 3.0 TTS ran on its default Plus tier. One shot each. Prices are the amounts charged for these exact requests. Times are wall-clock seconds for one request on October 18, 2026, from sending the request to receiving the finished file; all four models also stream audio while it is generated.
What to look for
Delivery. Listen for how each voice handles the turn from tired to joyful at "congratulations", and how much it changes pace across the last sentence.
Length and price. The same text produced between 12.3 and 16.8 seconds of audio. Qwen Audio 3.1 TTS and Gemini 3.8 Flash TTS bill the audio they generate, so a slower read costs more; Qwen Audio 3.0 TTS and StepAudio 3 TTS bill the characters you send.
Run the same test on EmpirioLabs
curl https://api.empiriolabs.ai/v1/audio/speech \
-H "Authorization: Bearer $EMPIRIOLABS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen-audio-3-1-tts",
"voice": "Ava_v3.1",
"input": "I opened the letter expecting another polite rejection. Instead, the first line said: congratulations, the job is yours. I read it three times before I let myself believe it.",
"instruction": "Start quiet and a little tired, then grow surprised and joyful by the last sentence.",
"response_format": "mp3"
}'Swap model and voice to run the same request against qwen-audio-3-0-tts, gemini-3-8-flash-tts (with style_prompt in place of instruction) or stepaudio-3-tts, or try them in the playground.
Frequently asked questions
Were the clips edited or retried?
No. Each clip is the first audio the model returned for this request, unedited.
Why are the clips different lengths?
Each model paces the same script in its own way, and the direction asks for a change of mood partway through. Length matters for price on Qwen Audio 3.1 TTS and Gemini 3.8 Flash TTS, which bill the generated audio.
Which one should I use?
Listen to the clips with your own script. The same request works for all four with only the model, voice and direction field changed.
