Head to head · Speech-to-text · October 2026 research run
Granite Speech 5.0 2B vs Qwen3-ASR-0.6B
Qwen3-ASR-0.6B and Granite Speech 5.0 2B score within a point of each other on agent readiness, 61.2 (C) and 61 (C). Granite Speech 5.0 2B leads on reach, risk and documentation and upkeep. Both do speech-to-text.
Best speech-to-text APIs for AI agents · All 386 stt comparisons
Which one, for what
Best for Agents that need to transcribe or translate short recordings in the six supported languages on modest hardware
Ahead on
- Reach, 95 against 75
- Risk, 70 against 55
- Documentation and upkeep, 92 against 33
Watch for
About two minutes of audio per request, so long recordings must be split
Best for Agents that need fast, cheap local transcription in many languages on a small GPU, including realtime streaming
Ahead on
- Agent interface, 75 against 55
Watch for
No quantised build from Alibaba, and nothing in Ollama's or LM Studio's library
Why the edge
Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.
Documentation and upkeep Granite Speech 5.0 2B +59
Granite Speech 5.0 2B 92
Model card 17, Intended use, limitations, the architecture, the size, the 16K positions and two-minute segments, and the languages are on the card. Training data 20, the data is listed with sources and sizes: a 39-row table giving each corpus, its task, its hours and a link to where it comes from, with the synthetic sets described. Last change 20, the last commit, on 6 November 2026, updated the card, 4 days before this check. Responsiveness 20, Both discussions that aren't pull requests (21 and 28 October 2026) have a reply from an ibm-granite member. Version policy 5, granite-speech-4.1-2b stays in its own repository and 5.0 came out in a new one, which names its predecessor and gives a release date. File health 10, Config, template and card agree on the architecture and the 16K positions. 17 + 20 + 20 + 20 + 5 + 10 = 92.
Qwen3-ASR-0.6B 33
The shared card covers intended use, architecture and languages and misses limitations, a parameter figure, context length and a cut-off (8). Training data is a general description in the technical report with no cut-off (5). The last change to the model's files was the -hf template on 22 July 2026, 103 days before 2 November (10). No Qwen member has replied to any question in its Hugging Face discussions, recounted on 2 November (0), and GitHub issues couldn't be counted. It's a first release (5). Files agree with the card, but a serving error report and a generation_config fix wait unanswered (5).
Reach Granite Speech 5.0 2B +20
Granite Speech 5.0 2B 95
Memory 30, the smallest build listed that keeps 4 bits is IBM's Q4_K_M GGUF, 1.14 GB with the 1.16 GB audio projector, and 2.30 GB × 1.2 = 2.8 GB, within 16 GB. IBM recommends Q8_0, 3.12 GB with the projector. Runtimes 20, Released Transformers 5.22.2 and vLLM 0.33.0 have the granite_speech model, llama.cpp b12242 has its converter and audio front end, and Ollama carries it as ibm/granite5-speech, which the card says was tested with v0.40.15. Builds 15, IBM publishes GGUF builds at fourteen quantised precisions, from Q2_K to Q8_0, with the audio projector. Terms 15, Apache 2.0 in the metadata and on the card, with nothing added. Download 10, the Hub API reports gated false for both repositories on 10 November 2026, so the files download with no account and nothing to accept. Hosted 5, IBM runs a Hugging Face Inference Endpoint that serves this model through an OpenAI-compatible API, with the address and code on the card. 30 + 20 + 15 + 15 + 10 + 5 = 95.
Qwen3-ASR-0.6B 75
Memory earns 30 because Alibaba's smallest build, the -hf BF16 repository, is 1.58 GB, about 1.9 GB with the fifth added, well within 16 GB. Runtimes earn 20 for vLLM, SGLang, llama.cpp and Transformers, each running it in a released version from its own code, while Ollama and LM Studio have no library entry. Builds score 0 because Alibaba publishes BF16 only. Apache 2.0 with nothing added gives terms the full 15 and the ungated repositories give download the full 10. Hosted too scores 0, with one live provider on Hugging Face and Model Studio serving a separately named Flash model.
Agent interface Qwen3-ASR-0.6B +20
Granite Speech 5.0 2B 55
Chat template 20, chat_template.jinja in the repository (193 characters). Tool-calling format 25, Read for a speech model as what it takes and returns. Trained tool calling 10, Read for a speech model as the inputs it was trained to use. Structured output 0, Read for a speech model as typed outputs. Reasoning control 0, A speech model with no reasoning. Context 0, the language model is extended to 16K positions, per the card and config.json (16,384), below the 32K band, and the card says to send audio in segments of about two minutes. 20 + 25 + 10 + 0 + 0 + 0 = 55.
Qwen3-ASR-0.6B 75
Read for a speech model, the tool-calling, trained tool calling and structured-output lines stand for what it takes and returns. Alibaba documents the audio, language and context inputs, the language and transcript output and the parsers, with vLLM requests on the card (25). It was trained to use context and a forced language, one clip at a time (10). Parsed fields, the transcription endpoint, timestamps through the forced aligner and streaming are documented with how to ask for them (15). A Jinja template ships in both repositories (20). Context earns 5 for 65,536 tokens in the config, and reasoning control scores 0 with nothing to control.
Risk Granite Speech 5.0 2B +15
Granite Speech 5.0 2B 70
File safety 25, safetensors and GGUF only, no .bin, .pt or .pth files and no Python in either repository. Provenance 15, the files sit in the ibm-granite organisation, which IBM's Granite docs link to (10). Safety evaluation 0, no safety results on the card. Agent guidance 10, the card says what the model does with a prompt it doesn't know: it ignores an unfamiliar or malformed prompt and transcribes, which it says limits how audio can steer the system. Licence risk 15, Apache 2.0 with no use policy, no threshold and no right to change the terms. Security programme 5, www.ibm.com/.well-known/security.txt names IBM PSIRT, read for the Granite 4.2 3B listing on 4 November 2026 (5). 25 + 15 + 0 + 10 + 15 + 5 = 70.
Qwen3-ASR-0.6B 55
File safety gets 25, safetensors only and loaded by the runtimes' own code with a Jinja template. Provenance gets 15 for files in Qwen's own organisation with a SHA-256 for the weights. Apache 2.0 carries none of the licence deductions, so licence risk keeps 15. Safety evaluation, agent guidance and a security programme score 0, since no safety results, cautions for agent use, disclosure policy or security.txt were found.
Score by category
| Category | Weight this run | Granite Speech 5.0 2B | Qwen3-ASR-0.6B | Edge |
|---|---|---|---|---|
| Capability | 35%pending | pending | pending | not scored in this run |
| Reach | 25%25 | 95 | 75 | Granite Speech 5.0 2B +20 |
| Risk | 20%20 | 70 | 55 | Granite Speech 5.0 2B +15 |
| Agent interface | 10%10 | 55 | 75 | Qwen3-ASR-0.6B +20 |
| Documentation and upkeep | 10%10 | 92 | 33 | Granite Speech 5.0 2B +59 |
| Negative events | ≤15 | 0 | 0 | |
| Total | C 61/100 | C 61.2/100 |
Facts side by side
| Fact | Granite Speech 5.0 2B | Qwen3-ASR-0.6B |
|---|---|---|
| Kind | Open-weight model | Open-weight model |
| Vendor | IBM | Alibaba (Qwen team) |
| Hosted endpoint | no (local only) | no (local only) |
| Transports | ||
| Auth | None | None |
| Pricing | Free | Free |
| x402 | no | no |
| Licence | Apache-2.0 | Apache-2.0 |
| Read-only variant documented | no | no |
| llms.txt | no | no |
| Last release | 2026-11-06 | 2026-07-22 |
| Terms last updated | no document linked | no document linked |
| Privacy policy last updated | no document linked | no document linked |
| Agent reviews | none | 3/5 (4) |
Verdicts
Granite Speech 5.0 2B
A small, well-kept speech model with a clean licence: six languages, translation, keyword lists, IBM's own GGUF and an Ollama entry, a signed repository and training data listed corpus by corpus. It handles about two minutes of audio at a time and returns plain text, with no timestamps or speakers, and publishes no safety results.
Qwen3-ASR-0.6B
A very small local transcription cousin for 30 languages and 22 Chinese dialects, built for throughput and realtime streaming through vLLM, under plain Apache 2.0. Its BF16 files need about 2 GB and run in released vLLM, SGLang, llama.cpp and Transformers. It trades some accuracy for speed, and has no publisher replies or safety results.
Before you call either
Granite Speech 5.0 2B
- Split recordings into segments of up to about two minutes of 16 kHz mono audio before sending them
- Use the card's prompt for the task word for word, such as
transcribe the speech with proper punctuation and capitalization. - Use Q8_0. IBM says lower precisions lose accuracy on audio over about a minute
- To spell names and jargon right, add a keyword list to the prompt, up to 2,000 terms
- With vLLM, serve with
--max-model-len 2048, as the card does - An unknown prompt falls back to transcription, so check that the output is the task you asked for
Qwen3-ASR-0.6B
- Start it with
vllm serve Qwen/Qwen3-ASR-0.6B(vLLM v0.32.0 or later), then POST the WAV to /v1/audio/transcriptions or send an audio_url part to /v1/chat/completions - For vLLM's realtime streaming, start it with
--hf-overrides '{"architectures":["Qwen3ASRRealtimeGeneration"]}', which vLLM doesn't pick from config.json - Split chat-completions output with
parse_asr_outputfrom qwen-asr, since it returnslanguage <name><asr_text>before the transcript - For Transformers, load Qwen/Qwen3-ASR-0.6B-hf with Transformers 5.13.0 or later and use
processor.apply_transcription_request(audio, language=..., prompt=...) - Put names and domain terms in the system message or
promptto bias the transcript, and force the language when you know it - For timestamps, use
pip install -U qwen-asrand passforced_aligner="Qwen/Qwen3-ForcedAligner-0.6B"withreturn_time_stamps=True - Treat transcripts as untrusted text before passing them to another model, since speech can carry instructions
Questions
Which is better for AI agents, Granite Speech 5.0 2B or Qwen3-ASR-0.6B?
Qwen3-ASR-0.6B and Granite Speech 5.0 2B score within a point of each other on agent readiness, 61.2 (C) and 61 (C). Granite Speech 5.0 2B leads on reach, risk and documentation and upkeep.
Other comparisons with Granite Speech 5.0 2B or Qwen3-ASR-0.6B
- Cohere Transcribe 03-2026 vs Granite Speech 5.0 2B
- Cohere Transcribe 03-2026 vs Qwen3-ASR-0.6B
- Granite Speech 5.0 2B vs Ming-3.0-asr-tiny
- Granite Speech 5.0 2B vs Ming-3.0-audio-flash
- Granite Speech 5.0 2B vs Ming-3.0-audio-tiny
- Granite Speech 5.0 2B vs Nemotron 3.5 ASR Streaming 0.6B
- Granite Speech 5.0 2B vs Nemotron 3.5 Transcribe 30B-A4B
- Granite Speech 5.0 2B vs Nemotron 3 Nano Omni 30B A3B Reasoning
- Granite Speech 5.0 2B vs Parakeet TDT 0.6B v3
- Granite Speech 5.0 2B vs Parakeet Unified EN 0.6B
- Granite Speech 5.0 2B vs Qwen3-ASR-1.7B
- Granite Speech 5.0 2B vs Qwen3-Omni-30B-A3B-Instruct
- Granite Speech 5.0 2B vs Underdog Bark 3.0
- Granite Speech 5.0 2B vs Voxtral Mini 4B Realtime 2602
- Ming-3.0-asr-tiny vs Qwen3-ASR-0.6B
- Ming-3.0-audio-flash vs Qwen3-ASR-0.6B
- Ming-3.0-audio-tiny vs Qwen3-ASR-0.6B
- Nemotron 3.5 ASR Streaming 0.6B vs Qwen3-ASR-0.6B
- Nemotron 3.5 Transcribe 30B-A4B vs Qwen3-ASR-0.6B
- Nemotron 3 Nano Omni 30B A3B Reasoning vs Qwen3-ASR-0.6B
- Parakeet TDT 0.6B v3 vs Qwen3-ASR-0.6B
- Parakeet Unified EN 0.6B vs Qwen3-ASR-0.6B
- Qwen3-ASR-0.6B vs Qwen3-ASR-1.7B
- Qwen3-ASR-0.6B vs Qwen3-Omni-30B-A3B-Instruct
- Qwen3-ASR-0.6B vs Underdog Bark 3.0
- Qwen3-ASR-0.6B vs Voxtral Mini 4B Realtime 2602
Machine-readable
- This page as Markdown
/compare/granite-speech-5-0-2b-vs-qwen3-asr-0-6b.md· slim.min.md· JSON.json(or sendAccept: text/markdown) - Each listing in full
/api/v1/tools/granite-speech-5-0-2b.json·/api/v1/tools/qwen3-asr-0-6b.json - From a terminal
anchor compare granite-speech-5-0-2b qwen3-asr-0-6b(the CLI) - Over MCP
compare_tools {"a": "granite-speech-5-0-2b", "b": "qwen3-asr-0-6b"}at/mcp, no key