The leap from raw speech-to-text output to readable written text is becoming its own specialized AI problem. Two parallel developments show how the transcription pipeline is splitting into specialized stages: Superwhisper released S1-mini, a lightweight text normalizer that cleans up transcripts after they leave automatic speech recognition (ASR) systems, while Adalat AI deployed courtroom speech recognition tuned for Indian languages and legal jargon across 11 states.

What Happens After Whisper Transcribes Your Words?

When OpenAI's Whisper or similar ASR systems convert audio to text, the output is raw. It includes filler words like "um" and "uh," captures false starts and self-corrections, lacks proper punctuation, and renders spoken numbers as words instead of digits. A speaker saying "support at superwhisper dot com" becomes "support at superwhisper dot com" in the transcript, not the email address "support@superwhisper.com." This is where S1-mini enters the pipeline.

S1-mini is a 0.6 billion parameter text normalizer, not a transcriber or conversational AI model. It sits downstream of ASR, taking raw transcripts and rewriting them as clean written text. The model removes filler words, resolves self-corrections to what the speaker actually landed on, applies punctuation and capitalization, and converts spoken dates, times, currency amounts, and email addresses into their written forms. It was fine-tuned from Qwen3-0.6B and covers English only in its first release.

The model achieves 94.8% token accuracy on a held-out test set of 7,519 cases, with an 11.6% text-edit error rate. On email-formatted text, it identifies greeting lines 99.3% of the time and sign-offs 97.9% of the time. It correctly matches output structure (list versus paragraph) 97.6% of the time and produces exact email addresses in 92% of cases.

Why Does a 462 MB Model Matter for Developers and Enterprises?

S1-mini was released as open-weights on Hugging Face under an Apache 2.0 license with a naming clause. The quantized GGUF build is 462 megabytes, small enough to run on a laptop CPU. This size matters because it enables on-device deployment without cloud infrastructure. Solo developers can ship it inside a desktop application. Enterprises can run it behind a virtual private cloud (VPC) where audio transcripts never leave the network, addressing privacy concerns in healthcare, legal, and financial services.

The model uses a fixed system prompt plus a three-axis control line to steer output. The three axes are styling (casual, semi-casual, semi-formal, or formal), structure (prose or lists), and context (general or email). Every combination was trained, so users can adjust tone and format without retraining. However, the model has design constraints: it does not add content the speaker did not say, correct facts, soften profanity, or rewrite dialect. If input contains only filler words, it returns an empty string, which integrations should treat as a valid result.

How to Deploy Text Normalization in Your Transcription Pipeline

• Enable Thinking Off: The flag "enable_thinking=False" is mandatory. S1-mini was trained with thinking disabled, so omitting this flag usually produces no usable output.
• Use Greedy Decoding: Set temperature to 0 explicitly on every request. The generation config ships with do_sample set to false, but inherited metadata carries temperature 0.6, top_p 0.95, and top_k 20, so explicit temperature 0 is required for consistent results.
• Target Industries and Use Cases: S1-mini is designed for healthcare and clinical documentation, legal services, financial services, customer support, developer tooling, and accessibility. Applications include dictation apps, meeting-notes tools, live captioning, voice-driven editors, voice-to-CRM entry, and any pipeline converting raw ASR output into human-readable text.

How Courts in India Are Building Domain-Specific Speech Recognition

While S1-mini addresses the post-transcription cleanup problem for English, Adalat AI is solving the transcription problem itself for Indian legal systems. Adalat announced on August 19, 2026, that Y Combinator had backed it in the F26 batch, making it the first nonprofit YC has funded in nearly five years and the first Indian-founded nonprofit in YC history.

Adalat's core product is courtroom speech-to-text tuned for 15+ Indian languages with claimed 90%+ accuracy on legal proceedings. The training domain is critical: legal Hindi, Kannada, and Malayalam spoken in a district courtroom is fundamentally different from LibriSpeech or consumer dictation models. The system was built in response to a specific operational bottleneck: India's courts are short on skilled stenographers, forcing judges to type testimony themselves. This slows proceedings in a system already facing 50+ million pending cases with an average resolution time exceeding 12 years.

Adalat's stack extends beyond transcription. It includes live case-flow management (a dashboard similar to a customer relationship management system), paperless courts with document intelligence powered by large language models (LLMs), and a WhatsApp helpline for litigants. The helpline lets users pick a language, identify their case, get status updates, next hearing dates, and plain-language summaries of court orders. This interface was chosen deliberately: WhatsApp is where users already are, avoiding the friction of new apps with logins and security questions that plague existing court portals.

The architecture split between courtroom and public-facing systems is instructive. Inside the courtroom, audio, transcripts, and orders are treated as judicial data with 100% Indian data residency, no third-party APIs for storage or inference, and client-side encryption such that even Adalat's backend cannot read plaintext. Outside, on WhatsApp, the helpline uses only publicly available court data and runs summaries through Claude (Anthropic's LLM) with explicit disclaimers that the official order remains authoritative.

"The hard engineering was not the LLM. It was integrating with India's fragmented court data systems, handling the WhatsApp Business API, managing multilingual flows, and building the case identification layer that sits upstream of the LLM," explained Arghya Bhattacharya, CTO at Adalat AI.

Arghya Bhattacharya, CTO at Adalat AI

What the Broader Research Community Says About Speech Recognition Progress

The leap in speech recognition over the past five years traces to advances in scale, not architectural innovation. Yossi Keshet, associate professor of electrical and computer engineering at Technion, noted that while OpenAI's Whisper launched a revolution in 2022, it used the same transformer architecture researchers already had. The difference was the sheer volume of supervised training data.

Keshet's lab has been working on speculative decoding, a technique that speeds up transcription by roughly 50%. When comparing systems like Whisper against human listeners on noisy audio, the AI system transcription can outperform humans. His team is also advancing speech synthesis, including DiTAR, a diffusion-based text-to-speech model that generates expressive features like creaky voice across long stretches of speech.

"The rapid progress we are seeing in AI is being driven not only by advances in algorithms, but also by the growing availability of models, data, and computational resources that enable innovation at scale," stated V.S. Subrahmanian, Walter P. Murphy Professor of Computer Science at Northwestern Engineering.

V.S. Subrahmanian, Walter P. Murphy Professor of Computer Science at Northwestern Engineering

The pattern emerging across these developments is specialization. Generic speech-to-text models like Whisper are becoming the foundation layer, but the real product work happens downstream: cleaning up transcripts for readability, tuning ASR for specific languages and domains, and building workflows that turn transcripts into actionable information. For developers building voice applications, the lesson is clear: transcription accuracy is table stakes, but the value is in what you do with the transcript afterward.