Slam 1

Slam-1 is AssemblyAI's groundbreaking Speech Language Model (SLM) that unifies large language model architecture with advanced automatic speech recognition (ASR) encoders to deliver superior speech-to-text transcription accuracy. Designed specifically for speech tasks, Slam-1 understands context and semantics at a deep level, enabling promptable and customizable transcription that adapts to specialized industry terminology and complex spoken content. This makes Slam-1 ideal for use cases in healthcare, legal, sales, and technical domains requiring precise and context-aware transcription.

‍

Technical Specifications

Performance Benchmarks

• Reduces missed entity rates by up to 66%, especially for names, medical, and technical terms.
• Decreases formatting errors by approximately 20%.
• Preferred by over 72% of end users in blind tests versus competing models.
• Achieves more reliable transcript quality in noisy and specialized contexts versus previous models.
• Delivers robustness against hallucinations through a multi-modal architecture that simultaneously processes audio and language.

Architecture Breakdown

Slam-1’s architecture distinctively merges a speech encoder with an adapter layer tuned to link acoustic features with a fixed large language model, enabling powerful semantic understanding. This multi-modal design surpasses traditional audio-to-text models by interpreting spoken content holistically, supporting accurate transcription and contextual reasoning. The approach leverages prompt engineering to customize transcription accuracy dynamically for industry-specific vocabularies and speech patterns.

API Pricing

• $0.00325 per min

Core Features & Capabilities

• Speech and Language Integration: Combines speech encoder and LLM for promptable and customizable transcription workflows.
• Fine-Tuning & Customization: Enables domain-specific adaptation through simple prompts without retraining.
• High Accuracy: Superior recognition of rare and domain-specific terms, improving downstream analytics and reducing manual review.
• Multi-Channel & Speaker Diarization: Supports complex audio streams with speaker separation and timestamps out of the box.
• Enterprise Ready: Designed to reduce post-processing effort and improve transcript quality in high-stakes industries like healthcare and legal.

‍

Code Sample

‍

Comparison with Other Models

VS AssemblyAI Universal: Slam-1 offers promptable, highly customizable transcription with superior entity recognition for specialized domains, while AssemblyAI Universal provides broader language support and lower latency for general transcription needs.

VS GPT-4.1 (audio transcription use): Slam-1 is specifically optimized for speech-to-text with multi-channel and speaker diarization features, unlike GPT-4.1, which focuses on general NLP tasks without native audio processing capabilities.