Hands-on DeepSeek's $0.06/hr ASR API: loses to a locally fine-tuned 1.7B model

AGI Hunt · wechat · 2026-11-14

DeepSeek opened up its app capabilities as standalone APIs, including a never-before-released speech recognition service at ¥0.39/hour (50% off launch pricing). The author benchmarked it on 95 clips of his real dictation (34 min) plus public sets (AISHELL, ASCEND, LibriSpeech, CommonVoice).

Key findings:

3.08% WER on his dictation — worse than even the unmodified local Qwen3-ASR 1.7B (2.57%), Qwen Flash (2.00%), Volcano (2.18%), ElevenLabs (2.17%)
Only 66% of English/technical terms correct in code-switched speech ("GitHub issue" became "help在艺术上"); the API accepts no hotwords, prompt, or context
Doubles sentence-final periods (291 vs 142 references), no CJK-Latin spacing, no streaming (0.84s median latency for short clips, 2.35s for long ones), 120-second max per request
Output-consistency analysis across benchmarks suggests the model is likely a retrained Qwen3-ASR 1.7B (agrees with it more than Qwen's own 1.7B vs 0.6B), though unconfirmed by DeepSeek

The workflow highlight: the author fine-tuned Qwen3-ASR 1.7B with LoRA on 829 self-recorded clips (4.8h), labeled by Claude Sonnet subagents, training in 5 min on a 4090. English-term accuracy rose 85%→94%, and the fine-tuned model hit 1.80% WER on held-out clips — beating every cloud provider. Verdict: skip DeepSeek's ASR until it adds hotword/context support.

Original post →

Hands-on DeepSeek's $0.06/hr ASR API: loses to a locally fine-tuned 1.7B model

More from coding & agent