DeepSeek ASR and TTS, tested: price, accuracy, latency

Pricing

• Speech to text: list price $0.116/hour, $0.058/hour during the 50%-off promotion, billed per second (partial seconds round up).
• Text to speech: list price $0.45 per 10K characters, $0.225 during the promotion. A Chinese character counts as 2; letters, digits and spaces count as 1.
• In practice that is about $22.5 per million English characters and $45 per million Chinese characters. See the speech model comparison for every provider side by side.

Speech-to-text accuracy

The same 50 clips (25 Mandarin, 25 English) through six engines. Lower is better.

Method: 25 Mandarin and 25 US English clips drawn at random from the Google FLEURS test set (about 9 minutes), converted to 16 kHz 16-bit mono WAV and loudness-normalised. Chinese is scored by character error rate (CER), English by word error rate (WER), with punctuation, case and script normalised. Latency is the median per request.

What we saw

• Chinese is the most consistent: its worst clip was 12% CER, versus 26% for Grok and 21% for Qwen. Most errors were transliterated foreign names.
• Very quiet recordings lose words: FLEURS has a dozen near-silent clips. Without loudness normalisation DeepSeek dropped the start or end of some English sentences and WER rose from 6.4% to 17.5%. Normalise volume first.
• Brand names drift: “BibiGPT” came back as “BB GPT”. There is no hotword parameter to fix it.
• Cantonese is recognised, but in the one clip we tried it came back as standard written Chinese rather than colloquial Cantonese characters.
• Fast: about 1 s per clip, on par with Qwen and StepFun and over twice as fast as Grok or Gemini.

Text-to-speech test

The same script (119 Chinese characters, 292 English characters, 83 Japanese characters) in both voices.

What we saw

• Quick to start: first audio in 0.8–1.3 s and synthesis at about 5× realtime, good for streaming playback.
• Numbers are read sensibly in Chinese. In English, “$0.06” is read as “zero point oh six dollars”.
• Only two voices, Stella (female) and Tide (male), with no speed, emotion or voice-cloning controls. Use MiniMax or ElevenLabs for that.
• Content review applies: rejected text returns HTTP 400.

Limits at a glance

Using it in BibiGPT

• Narration for image videos and document explainers: pick “DeepSeek TTS” as the voice engine. Pro members need no API key.
• For long-video subtitles keep Auto: DeepSeek ASR has no timestamps and a 120 s cap, so BibiGPT has not added it as a subtitle engine yet.

Calling the API

Both endpoints use your regular DeepSeek API key, the same one as the chat models. ASR docs · TTS docs

curl https://api.deepseek.com/v1/tools/asr \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -F file="@speech.wav"
# → {"request_id":"…","text":"…","duration_ms":3200}

# MP3 / M4A 先转成 16 kHz 单声道 WAV，切成 ≤120 秒一段
ffmpeg -i in.mp3 -ac 1 -ar 16000 -sample_fmt s16 \
  -f segment -segment_time 110 part_%03d.wav

curl https://api.deepseek.com/v1/tools/tts \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"text":"你好，DeepSeek。","voice_id":"tide","format":"wav"}' \
  --output speech.wav