How ISSEN delivers AI language tutoring at scale with Sprag

ISSEN (YC F24) is redefining what a language learning app can be. Instead of preset prompts and gamified drills, ISSEN gives learners free-flowing, human-like voice conversations with an AI tutor that adapts to their level, interests, and goals across more than 50 languages. Whether a learner is practicing Japanese, Spanish, or Gujarati, ISSEN recreates the experience of a dedicated human tutor.

That experience depends on a fast, accurate voice pipeline. A tutor that mishears a learner, lags, or struggles across languages breaks the immersion instantly. To power the speech-to-text (STT) layer behind those conversations, ISSEN turned to Sprag, the voice AI inference provider lowering latency and cost across the voice stack.

Built on Sprag, ISSEN's transcription runs on a high-quality multimodal model with low latency, broad language coverage, and a cost profile the two teams continue to optimize together.

#The challenge: replacing a core model without regressing quality

ISSEN relied on Gemini Flash 2.0 because it offered strong pricing and performance across many languages, which is essential when you're supporting users that speak more than 50 of them. When that model was abruptly deprecated, ISSEN needed a replacement immediately, and it could not be a step down.

The bar was high on three fronts:

• Multilingual accuracy. Speech recognition is one of the biggest sources of friction in voice AI. ISSEN's learners switch languages, accents, and proficiency levels constantly, and any weakness in transcription is felt directly in the product.
• Latency. Realtime conversation means transcription cannot add perceptible delay.
• Cost at scale. With a steady, always-on user base across many time zones, small per-request differences compound quickly.

#The solution: transcription served and optimized on Sprag

ISSEN turned to Sprag to run a large multimodal transcription model in production without the usual penalties in cost, latency, or scale. Key elements included:

• Higher-quality transcription. ISSEN runs a large multimodal model as its transcriber rather than the smaller, dedicated ASR model, which keeps accuracy high across all of its supported languages. Sprag's serving stack makes the larger model practical to run at production latency and cost.
• Low-latency serving. Sprag serves the model on a vLLM gateway, with context caching and prefix-aware routing enabled by default and STT billed on input audio duration. In production, ISSEN sees less than 300ms transcription latency, fast enough to keep realtime conversations feeling natural.
• Throughput on demand. As ISSEN rolled out to its full user base, Sprag raised throughput limits the same day and provisioned headroom ahead of further growth, so the deployment was never gated by capacity.
• Joint cost optimization. Because transcription cost is driven mostly by audio duration, Sprag optimized it alongside ISSEN's team to reduce costs, matching the model's native sample rate, trimming silence to cut wasted tokens, quantizing the model to shrink its memory footprint, and running dedicated instances tuned to ISSEN's transcription-only workload.

"I've been working with multilingual voice AI for years, testing dozens of transcribers and companies, and Qwen-Omni on Sprag has been the highest performing model at very low latency. We have difficult evals that test noisy, multilingual code-switched audio, with a variety of accents. Most models like Deepgram, 11labs scribe, etc land at around 70% on our evals, while Sprag has reached 90%, with a latency under 300ms."

• Mariano Sorgente, Founder and CEO of ISSEN

#The impact

By moving their STT to Sprag, ISSEN replaced a deprecated model with a stronger one, and did it quickly:

• A fast, seamless cutover. ISSEN migrated its transcription workload to Sprag and rolled it out to its full user base without degrading its learner experience.
• Realtime latency. Less than 300ms transcription latency keeps voice conversations feeling natural.
• A clear path on cost. Silence trimming and a quantized model variant meaningfully reduced per-request cost, with further optimization underway.
• Room to grow. Throughput was provisioned well beyond ISSEN's current peak, so scaling further does not require re-architecting.

"Sprag has been great to work with and extremely helpful with deployment, custom deals and support. I look forward to working with them in the future."

• Mariano Sorgente, Founder and CEO of ISSEN

Want to lower latency and cost across your voice AI pipeline? Learn more about Sprag.