Skip to content

ADR-v2-141 — Chinese P1 stays on the existing Whisper engine; alternatives must win by measurement

Status: Proposed (2026-09-20)
Context: Chinese ASR model selection; CPU-first/offline architecture; SttEngine extensibility.

Context

There are now several strong open ASR families with Chinese coverage, including Qwen3-ASR and FunASR checkpoints such as SenseVoiceSmall and Paraformer-zh.

Adding one immediately is tempting, but YazSes is not only an ASR benchmark harness. Its recognizer must fit:

  • the CPU-first hold-to-talk latency budget;
  • Linux/macOS/Windows packaging;
  • streaming and batch paths;
  • word timestamps;
  • air-gapped provisioning;
  • local/offline operation;
  • current memory/install expectations;
  • model licensing;
  • the existing SttEngine seam.

The repository already supports multilingual Whisper through faster-whisper and has a measured Chinese script-normalization path.

Decision

For the first first-class Mandarin release, use the existing:

[stt]
engine = "faster-whisper"
model = "small"
language = "zh"

with the existing Han-script decorator.

small is the supported baseline candidate, not a permanent declaration that Whisper is the best Chinese recognizer.

Benchmark alternative models through the existing SttEngine protocol. A new runtime is optional and lazy-loaded; it does not become a base dependency merely because it has better upstream benchmark numbers.

Evaluation priority

  1. Same-runtime checkpoints: large-v3-turbo, large-v3 (Whisper code + weights: MIT).
  2. Broad dialect candidate: Qwen3-ASR-0.6B (official model card: Apache-2.0).
  3. Chinese streaming candidate: Paraformer-zh / Paraformer-zh-streaming, pinned to an exact checkpoint/revision whose model weights explicitly declare Apache-2.0.
  4. SenseVoiceSmall as a research-only comparison while its weights use the custom FunASR Model Open Source License; it is not a permissive-license production candidate.
  5. Larger models only where target hardware/use case justifies them.

A toolkit's source-code license is not sufficient evidence for a model checkpoint. Every production candidate records the exact weight repository, immutable revision/commit and weight-license source. YazSes prefers MIT/Apache-2.0/BSD-style model-weight licenses for recommended profiles; a custom model license requires explicit maintainer/legal review before bundling or recommending the weights.

Promotion requirement

A new engine becomes a recommended profile only if YazSes measurements demonstrate a material product benefit, such as:

  • =20% relative CER reduction, or

  • materially better validated dialect coverage, or

  • =30% latency/core-cost reduction at non-inferior quality, or

  • a supported language capability the baseline lacks,

while also passing packaging, timestamp, streaming, offline and license gates. The license gate requires an auditable weight license for the exact revision; permissive licenses are the default for recommended profiles.

Consequences

Positive

  • P1 is mostly orchestration/test work rather than a second ML stack.
  • English installation size and runtime stay unchanged.
  • Chinese support can ship independently from research into specialized models.
  • Model changes remain replaceable behind one interface.
  • Upstream model hype does not silently redefine YazSes’ hardware contract.

Costs

  • P1 may not deliver the absolute best achievable Mandarin CER.
  • The project must build and maintain a reproducible benchmark harness before choosing a specialized model.
  • Supporting a future Chinese engine still requires an adapter and package matrix.

Alternatives considered

Adopt Qwen3-ASR-0.6B immediately

Rejected for P1. Its broad Chinese dialect support is compelling, but the official runtime/deployment path introduces a substantially different stack, and interactive CPU behavior has not been measured in YazSes.

Adopt SenseVoiceSmall immediately

Rejected for P1 and not a permissive-license production candidate at present. It is promising for CPU Chinese/Cantonese and may be benchmarked for research comparison, but its current weights point to the custom FunASR Model Open Source License rather than MIT/Apache/BSD. Any future production use requires explicit review of the exact weight revision in addition to adapter, timestamp, packaging and comparative measurements.

Use large-v3 as the default Chinese model

Rejected as an unmeasured default for the CPU-first product. Existing microbench evidence suggests better CER, but its resource/long-form tradeoffs are substantial and the repository has already observed large-v3 conditioning instability on other corpora.

Auto-select a model by available RAM/CPU

Deferred. First establish validated presets and capability data. Hidden automatic model changes make results harder to reproduce.

Required evidence before superseding this ADR

A superseding ADR must include:

  • exact model repository, immutable revision/commit and model-weight license source;
  • corpus/split/normalization;
  • CER and error analysis;
  • RTF/latency/RSS/core-seconds;
  • startup/cache/offline behavior;
  • SttEngine method coverage;
  • package impact on each supported OS;
  • streaming and meeting behavior;
  • comparison against the currently supported Chinese baseline.