Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchCohere releases Transcribe, a 2B parameter open-source ASR model supporting 14 languages. It achieves 5.42 WER and processes 525 minutes of audio per minute on a consumer GPU. Free API access included. The speech recognition market just got commoditized.
AI, startups & growth insights. No spam.
Cohere just commoditized enterprise speech recognition. Its new Transcribe model — 2 billion parameters, Apache 2.0 licensed, deployable on a consumer GPU — achieves a 5.42 Word Error Rate on the Open ASR Leaderboard, outperforming ElevenLabs Scribe v2, OpenAI Whisper Large v3, and every other publicly benchmarked model in its class. The closed-source transcription market didn't see this coming.
Released on March 26, 2026, Cohere Transcribe is the company's first open-source audio model. It supports 14 languages, processes audio at roughly 3x the throughput of comparably-sized competitors, and is available immediately — both through a rate-limited free API on Cohere's dashboard and as a fully self-hostable model under a permissive open-source license. For enterprises paying per-minute rates to ElevenLabs, Deepgram, or AssemblyAI, the math just changed dramatically.
Cohere Transcribe is a 2-billion-parameter automatic speech recognition model built on an encoder-decoder transformer architecture. The encoder uses a Fast-Conformer design — a proven architecture for audio that scales efficiently without the quadratic attention costs of vanilla transformers. More than 90% of the model's parameters live in the encoder, with a deliberately lightweight decoder that minimizes the autoregressive inference bottleneck. That's an intentional design choice: most of the heavy lifting happens once, not token-by-token.
The model was trained on 500,000 hours of curated audio-transcript pairs, augmented with synthetic data at signal-to-noise ratios between 0 and 30 dB. Audio is preprocessed at 16kHz. A 16,000-token multilingual BPE tokenizer with byte fallback handles the vocabulary — trained specifically on in-distribution data rather than borrowed from a text LLM, which matters for rare words, accents, and named entities.
Licensing is Apache 2.0. That means commercial use, modification, and redistribution are all permitted without restriction. There are no per-seat fees, no usage caps on self-hosted deployments, no vendor lock-in clauses. The weights are available on Hugging Face. You can run it today.
Access comes in three tiers. The Hugging Face Space offers a free demo. The Cohere dashboard provides rate-limited free API access for developers building and prototyping. Enterprise teams needing production throughput without rate limits can provision dedicated instances through Cohere's Model Vault — billed by the hour with discounts for longer commitments.
The headline number is 5.42 average Word Error Rate on the Open ASR Leaderboard — the community-standard benchmark for English speech recognition across eight diverse test sets: AMI meeting transcriptions, Earnings22 financial calls, Gigaspeech podcast audio, LibriSpeech Clean and Other, SPGISpeech, Tedlium lectures, and Voxpopuli parliamentary speech.
That 5.42 WER is a #1 ranking. The next closest open model is Zoom Scribe v1 at 5.47 — a gap of 0.05 percentage points. IBM Granite 4.0 Speech 1B comes in third at 5.52. OpenAI's Whisper Large v3, still the most widely deployed open-source ASR model in production today, scores 7.44 — a 37% worse WER than Cohere Transcribe on the same benchmark suite.
On individual test sets, Transcribe wins decisively on LibriSpeech Clean (1.25 WER) and LibriSpeech Other (2.37 WER) — the clean-speech and noisy-speech splits that best reflect real deployment conditions. Its AMI score of 8.15 beats IBM's 8.44 and ElevenLabs' 11.86 on meeting transcription, which is arguably the highest-value commercial use case.
Speed is the second critical metric. Cohere reports 3x higher offline throughput than similarly-sized competitors. Production inference runs through vLLM with custom encoder-decoder support — a pipeline that includes fine-grained concurrent execution of variable-length encoder requests, packed representation for decoder inputs, FlashAttention-based decoding, and dynamic KV-cache management. The result is up to 2x additional throughput improvement on top of the baseline model speed. For a batch transcription workload — meeting recordings, call center audio, podcast archives — that throughput advantage compounds directly into infrastructure cost savings.
Human evaluation data reinforces the automated benchmarks. In head-to-head comparisons, Transcribe achieves preference scores above 50% versus open-source alternatives across accuracy, coherence, usability, and named entity recognition — the latter being a persistent weakness in most ASR models when handling product names, company names, and technical jargon.
OpenAI Whisper Large v3 remains the baseline comparison for any serious ASR model. Whisper popularized the concept of a general-purpose, open-weight speech recognition model, and its 1.5-billion-parameter architecture has been the default choice for self-hosted transcription since 2022. Transcribe beats it on WER by 27% (5.42 vs. 7.44), processes audio faster, and handles multilingual inputs more reliably — particularly on non-European languages like Arabic, Vietnamese, and Korean where Whisper's training data was thinner. For any team currently running Whisper in production, the upgrade path is clear.
ElevenLabs Scribe v2 is the most direct closed-source competitor. It scores 5.83 WER on the Open ASR Leaderboard — 7% worse than Transcribe. ElevenLabs doesn't publish Scribe's architecture, weights, or pricing tiers in detail. Cohere Transcribe's human evaluation data shows Transcribe is preferred in multilingual settings specifically, which is where Scribe had built its reputation. The combination of better benchmarks and open weights makes the comparison unfavorable for ElevenLabs.
Deepgram and AssemblyAI don't participate publicly in the Open ASR Leaderboard, making direct WER comparison impossible. Both operate on per-minute pricing models — Deepgram Nova-3 starts around $0.0043/minute for pre-recorded audio; AssemblyAI's Universal-2 runs $0.006/minute. For a team transcribing 100,000 minutes per month, that's $430–$600 per month before any volume discounts. A self-hosted Transcribe deployment on commodity GPU hardware eliminates that recurring cost entirely, with only compute overhead remaining.
NVIDIA Canary Qwen 2.5B scores 5.63 on the leaderboard — 4% worse than Transcribe despite comparable parameter counts. IBM Granite 4.0 Speech 1B is 2% worse at 5.52, though it achieves this with half the parameters, making it the more efficient option if raw compute cost is the primary constraint.
The honest competitive summary: Cohere Transcribe is #1 on every major English benchmark, open-source under Apache 2.0, and self-hostable on consumer hardware. There is no closed-source competitor that can match all three criteria simultaneously.
Cohere's decision to release Transcribe as open-source under Apache 2.0 is a calculated competitive move — and it reflects a broader pattern playing out across the AI industry in 2026.
The company's core business is enterprise AI deployment through its North platform. North sells large organizations on governed, secure AI infrastructure with compliance controls, audit trails, and SLA-backed uptime. The revenue model depends on enterprises choosing Cohere's managed platform over competitors' clouds or DIY deployments. Open-sourcing individual models — particularly specialized ones like ASR — serves that strategy in two ways.
First, it drives developer adoption. A developer who builds a transcription pipeline on Cohere Transcribe has a natural path to Cohere's API for production deployment, then to Model Vault for enterprise scale, then to North for full governance. The open weights are the top of a commercial funnel, not a departure from it.
Second, open-sourcing commoditizes a capability that competitors charge for. If ElevenLabs Scribe v2 is a meaningful revenue line for ElevenLabs, a free, better-performing open alternative directly attacks that revenue. Cohere doesn't need to win the transcription API market — it needs to ensure the transcription API market doesn't fund a competitor's growth into Cohere's core enterprise territory.
This playbook is identical to what Meta executed with Llama, what Mistral executed with its base models, and what Google has done selectively with Gemma. The model is free. The infrastructure, governance, and enterprise SLA are not.
Cohere Transcribe supports English, German, French, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Arabic, Vietnamese, Chinese (Mandarin), Japanese, and Korean. That's 14 languages spanning Latin, Semitic, Southeast Asian, and East Asian scripts — handled by a single 16,000-token multilingual BPE tokenizer with byte fallback.
The byte fallback is significant. It means the model can handle characters outside its explicit vocabulary without failing — critical for mixed-script inputs, unusual proper nouns, and code-switching at the word level. The tokenizer was trained on in-distribution audio transcription data rather than repurposed from a general text LLM, which gives it better calibration on the specific vocabulary distributions that appear in speech — including hesitations, filler words, and spoken number formats.
The HuggingFace blog post shows multilingual human evaluation data where Transcribe achieves strong preference scores against open-source alternatives across non-English languages. The model requires a language tag for optimal performance — it expects monolingual audio per inference call. Code-switching support (switching languages mid-sentence) is listed as a known limitation, though byte fallback provides partial resilience.
For the European languages (German, French, Italian, Spanish, Portuguese, Dutch, Polish, Greek), the model benefits from overlapping training data quality. Arabic and the Asian languages (Vietnamese, Chinese, Japanese, Korean) represent the more challenging cases — and Cohere's explicit inclusion of these in human evaluation rather than hiding them in asterisked footnotes is worth noting. Most ASR model releases bury multilingual performance in supplementary materials.
The 2B parameter size is not an accident. It sits in a deliberate sweet spot: large enough to outperform all sub-2B competitors on accuracy, small enough to run on a single consumer GPU without quantization tricks.
The Fast-Conformer encoder architecture scales with linear attention rather than quadratic — meaning longer audio sequences don't cause memory explosions the way they would with standard transformer attention. This is what enables processing long-form audio (an hour-long meeting recording, a two-hour podcast) without chunking it into small windows and stitching transcripts back together. Chunking introduces stitching artifacts and misses words at boundaries. Full-sequence processing avoids both.
The vLLM integration is what makes production throughput viable. vLLM's continuous batching allows the model to process multiple audio files concurrently with variable lengths — a batch of thirty 90-second clips alongside a single 45-minute recording, all in the same inference pass. The custom encoder-decoder support Cohere contributed to vLLM includes packed representation for decoder inputs (eliminating padding overhead between sequences) and FlashAttention-based decoding (reducing memory bandwidth requirements per step). Together, these optimizations yield up to 2x better throughput than a naive HuggingFace inference setup on identical hardware.
A single RTX 4090 — a consumer GPU available for under $2,000 — is sufficient for self-hosted deployment. For batch workloads (asynchronous transcription of recorded audio), a single 4090 handles substantial throughput. Real-time streaming transcription of multiple concurrent audio streams would require more GPU memory or multiple cards, but the baseline accessibility bar is genuinely low.
For teams that don't want to manage their own GPU infrastructure, Cohere provides two enterprise-grade access paths.
Model Vault offers dedicated model instances provisioned through the Cohere dashboard. Unlike shared API endpoints — where rate limits and multi-tenant latency variance affect production reliability — Model Vault allocates compute exclusively to your workload. Pricing is hour-instance based, with discounts available for longer-term commitments (30-day, 90-day, annual). There are no per-minute transcription fees; cost scales with compute time rather than audio volume.
This pricing structure is particularly favorable for high-volume workloads. A call center processing 500,000 minutes of audio per month would pay significant per-minute rates with Deepgram or AssemblyAI. On a dedicated Model Vault instance running continuously, the cost becomes a fixed infrastructure line — predictable, auditable, and significantly lower at scale.
Cohere North is the broader enterprise platform that wraps Model Vault with governance tooling: role-based access controls, audit logging, compliance certifications, and integrations with enterprise identity providers. For organizations in regulated industries — financial services, healthcare, legal — North provides the operational envelope that makes AI deployment defensible to compliance and security teams. Transcribe's accuracy on financial earnings calls (10.84 WER on Earnings22, second-best in class) makes it directly relevant to those sectors.
Cohere also provides a documentation portal covering rate limits, integration patterns, and API reference. The free tier on dashboard.cohere.com is rate-limited but fully functional — suitable for development, testing, and low-volume production use before committing to a paid tier.
The speech recognition market has followed the same arc as every AI capability before it: proprietary → open-source parity → open-source superiority → commoditization. We arrived at that last stage today.
Three years ago, enterprise ASR meant Google Cloud Speech-to-Text or AWS Transcribe — black boxes with per-minute billing and limited customization. OpenAI Whisper in late 2022 demonstrated that open weights could match commercial accuracy. ElevenLabs Scribe, Deepgram Nova, and AssemblyAI Universal-2 built proprietary moats on top of improved architectures and specialized fine-tuning. Those moats just got flattened.
When an open-source model outperforms the best closed alternatives on every standard benchmark, runs on commodity hardware, and carries no licensing restrictions, the closed-source incumbents face an existential pricing question. They cannot charge $0.005/minute for a capability that developers can deploy for the cost of GPU electricity. They must either move upmarket (real-time streaming, speaker diarization, domain-specific fine-tuning, compliance infrastructure) or compete on integrations and developer experience rather than raw model accuracy.
This is not hypothetical — it's the pattern that played out in text generation after Llama, in image generation after Stable Diffusion, and in code generation after CodeLlama. The model becomes a commodity. The value migrates to the infrastructure, the workflow, and the integration layer.
For Cohere, commoditizing transcription is safe — the company doesn't derive meaningful revenue from a transcription API. For ElevenLabs, Deepgram, and AssemblyAI, it's a direct threat to a product line. The response will be instructive.
The practical implications for developers are immediate. Any application that currently uses Whisper can upgrade to Transcribe and expect meaningful accuracy improvements, particularly on meeting audio, accented speech, and non-English content. The HuggingFace weights are available now. The vLLM integration means the model slots into existing inference infrastructure with minimal friction.
For new applications, Transcribe opens several use cases that were previously cost-prohibitive at scale. Podcast transcription at volume — even a medium-sized podcast network with 10,000 hours of back catalog — becomes a one-time compute job rather than an ongoing API cost. Call center analytics, where every customer call is transcribed and analyzed, shifts from a per-minute billing model to fixed infrastructure. Meeting intelligence products (real-time notes, action items, summaries) can run entirely on-premise for organizations with data residency requirements.
The known limitations are worth flagging honestly. The model requires explicit language tags — if you're processing audio where the language is unknown, you'll need a language detection step upstream. It's also eager to transcribe non-speech sounds, which means pauses, background noise, and music sections may produce hallucinated transcriptions without a Voice Activity Detection gate in front of it. Both are solvable engineering problems, not fundamental accuracy issues, but they require deliberate pipeline design.
Cohere's documentation covers both limitations with mitigation guidance. The customizable punctuation prompting (inherited from the Canary architecture) allows downstream formatting control — useful for applications where transcript formatting consistency matters (legal transcription, medical dictation, closed captioning).
The free rate-limited API tier makes the initial integration path zero-cost. Build your pipeline against the API, validate accuracy on your specific audio domain, then make the self-hosting versus Model Vault decision based on your volume and infrastructure preferences.
Cohere Transcribe is the most accurate publicly benchmarked ASR model available today. It's open-source. It runs on consumer hardware. It supports 14 languages. It's free to use within rate limits and self-hostable without restriction.
The release marks a clear inflection point for the speech recognition market. The accuracy gap between open-source and closed-source ASR — which justified per-minute API pricing for the past three years — has closed. Developers no longer have to choose between accuracy and cost control. Enterprises no longer have to choose between data sovereignty and state-of-the-art performance.
Cohere's positioning is smart: open-source the model, monetize the infrastructure. The North platform and Model Vault provide the enterprise operating environment that most organizations need but few want to build themselves. The free weights drive adoption and build ecosystem momentum. The transcription API market becomes a loss leader for a much larger enterprise infrastructure play.
What happens next depends on how incumbents respond. ElevenLabs, Deepgram, and AssemblyAI will need to differentiate on real-time streaming latency, domain-specific fine-tuning, diarization accuracy, and workflow integrations — capabilities where pure WER benchmarks don't capture the full picture. There's still a market for managed transcription services. It's just a harder market to defend than it was yesterday.
For developers building today: download the weights, read the HuggingFace technical blog, and test against your specific audio domain before committing to any inference architecture. The accuracy numbers are compelling. The production engineering requirements — VAD preprocessing, language detection, vLLM setup — are manageable. The cost case is unambiguous.
Speech recognition is now a solved problem for the majority of production use cases. The next battleground is real-time, the next frontier is multilingual code-switching, and the next capability to commoditize is probably speaker diarization. Cohere fired the starting gun. The rest of the field has to respond.
Sources: Cohere Transcribe Technical Blog — HuggingFace | TechCrunch coverage | Open ASR Leaderboard | Cohere Dashboard
NVIDIA announces NemoClaw at GTC 2026 — an open-source enterprise AI agent platform that integrates with the Vera Rubin hardware stack, positioning NVIDIA as the full-stack agentic AI infrastructure provider.
Threat actor TeamPCP compromised LiteLLM v1.82.7 and v1.82.8 on PyPI through a stolen credential attack via a poisoned Trivy GitHub Action, deploying a malicious .pth file that silently exfiltrated API keys and auth tokens from any Python environment that imported the package.
Anthropic previewed auto mode for Claude Code on March 24-25, 2026 — a classifier-based system that lets the AI agent act autonomously on safe operations while redirecting risky ones, avoiding mass file deletions, sensitive data leaks, and malicious code execution.