SMSpeech Model Match
Home / Guides / Meeting transcription
Workload guide · 2026

Best local speech-to-text for meetings and speaker attribution

Meeting transcription is not just an ASR accuracy problem. A useful local stack also needs speaker attribution, stable long-form behavior, sensible timestamps and — for live calls — latency that does not drift as the session grows.

Short answer: benchmark VibeVoice-ASR first when you want one model to produce speaker-attributed, timestamped long-form transcripts. For a mature modular workflow, Whisper or faster-whisper plus WhisperX remains attractive. For live meetings, test VibeVoice-ASR-Streaming and Nemotron rather than assuming a chunked batch model will behave like a streaming architecture.

What a meeting stack actually has to solve

A raw transcript answers only “what was said.” Meeting notes usually need at least four additional properties: who said it, when it was said, whether speaker identity stays consistent over a long session, and whether names or domain vocabulary survive noisy multi-speaker audio. If the meeting is transcribed live, partial-result stability and end-of-turn latency matter as well.

This is why a model with slightly better word error rate can still be a worse meeting choice. A separate diarization stage, forced aligner and post-processing pipeline add latency, memory use and failure modes. Conversely, a model that emits speaker labels natively may still need correction if speaker boundaries or names are unreliable on your room acoustics.

VibeVoice-ASR: the most integrated long-form candidate

Microsoft's long-form VibeVoice-ASR is designed for recordings up to 60 minutes in a single pass and generates a structured transcript containing speaker identity, start/end timestamps and content. Its current Hugging Face card lists 51 languages and an MIT license. For interviews, podcasts and meetings, that is unusually close to the output schema people actually want.

The main cost is hardware. Although the family is marketed as VibeVoice-ASR-7B, the current Hugging Face artifact reports 9B total parameters. Treat it as a heavy model and test memory headroom, generation speed and long-session stability on the machine that will actually run it.

Whisper + faster-whisper + WhisperX: the modular baseline

Whisper itself does not provide native speaker diarization. A common local architecture is Whisper or faster-whisper for transcription, followed by WhisperX-style forced alignment and an optional diarization component. The advantage is modularity: you can change the Whisper size or runtime without replacing the whole pipeline, and the ecosystem is mature.

The disadvantage is exactly the same modularity. Each component has its own model/runtime requirements, error modes and license considerations. Speaker diarization should be evaluated separately from ASR accuracy, and alignment quality can change by language.

Live meetings: VibeVoice Streaming versus Nemotron

VibeVoice-ASR-Streaming continuously emits speaker-attributed “who said what” output and supports customized hotwords. The current 1.5B-named checkpoint reports 3B total parameters and ten languages. That makes it a direct candidate for live multi-speaker transcription rather than merely a faster batch recognizer.

Nemotron 3.5 ASR Streaming 0.6B is much smaller and is explicitly designed around low-latency streaming, but it does not natively solve speaker diarization. If the application only needs a live transcript, that may be the cleaner architecture. If speaker attribution is a hard requirement, the extra pipeline stage changes the comparison.

CPU-only meetings: VibeVoice-ASR-BitNet is worth a benchmark

The BitNet variant compresses the VibeVoice ASR runtime to about 1.58 GB and its model card reports real-time throughput with three or more CPU threads on the vendor's tested system. It covers seven languages. That makes it unusually interesting for offline meeting appliances and CPU-only desktops.

Do not interpret vendor RTF as a universal guarantee, and do not confuse real-time throughput with native incremental streaming. Test a complete 30–60 minute recording, not just a 20-second clip, and watch both memory growth and speaker consistency.

A practical meeting benchmark

  1. Use at least one 30–60 minute recording with real interruptions and overlapping speech.
  2. Include at least three speakers, similar voices if that occurs in production, and several names or technical terms.
  3. Score transcription errors and speaker-attribution errors separately.
  4. Measure whether speaker IDs remain stable after long silences and topic changes.
  5. For live systems, record first partial text, stable-final latency and revision churn.
  6. Measure total pipeline memory, not just the ASR checkpoint.

Recommended starting paths

One-pass long-form meeting transcript: VibeVoice-ASR. Mature modular workflow: faster-whisper + WhisperX / dedicated diarization. Live speaker-attributed meeting: VibeVoice-ASR-Streaming. Small native streaming ASR with separate diarization: Nemotron 3.5. CPU-only supported-language test: VibeVoice-ASR-BitNet.

For live-only latency, see best local STT for streaming. For subtitle timing and alignment, see best local STT for subtitles and timestamps. If you are deciding whether to keep Whisper but change the runtime, see Whisper vs faster-whisper.

Related guides

Primary sources

Features and vendor benchmark claims are source-derived; ranking language on this page is editorial guidance. Re-test on your own audio before production.

Reviewed against primary sources on September 13, 2026. Model behavior, runtime support, language coverage and licenses can change; re-check the linked primary source before production use.