whisper-large-v3-turbo

Whisper large-v3-turbo — 30 s TFLite (LiteRT), quantized

Quantized TFLite (LiteRT) exports of openai/whisper-large-v3-turbo: 30 s fixed-window graphs split into encode and decode signatures, matching the graph interface of litert-community/whisper-tiny and litert-community/whisper-base — drop-in for pipelines built against those graphs (with the 128-mel frontend below).

Exported by the LiteRT-LM-Unity project from the openai checkpoint (transformers TFWhisperForConditionalGeneration → two-signature 30 s graph), then post-training quantized. The f32 source export (~3.2 GB) is retained by the producing project and available on request.

Files

Integration notes (128-mel / vocab 51866)

• 128 mel bins — the large-v3 family needs a 128-mel log-spectrogram frontend, not the 80-mel frontend used by tiny/base/medium. An 80-mel input silently produces wrong output.
• Vocab 51866 — special token ids shift vs the 51865-vocab models: <|transcribe|> = 50360, <|notimestamps|> = 50364 (language tokens shift accordingly, e.g. <|ko|>/<|en|> differ from the tiny/base ids).
• Two signatures: encode (128-mel spectrogram, 30 s window → encoder hidden states) and decode (token ids + encoder states → logits). The decoder is a fixed-length full re-run per step (no KV cache), matching the tiny/base graph interface.
• Tokenizer: use tokenizer.json from openai/whisper-large-v3-turbo.

Quantization

Quantizer: ai-edge-quantizer 0.8.0, post-training dynamic-range.

• i8: int8 weights channelwise, fp32 activations (dynamic_wi8_afp32 scheme).
• i4 ("mixD"): dynamic_wi4b64_afp32 (int4 weights, blockwise-64, fp16 scales) with int8 overrides on the token-embedding/logits table. Pure full-scope int4 corrupts Korean output on the large family and was rejected. int4 channelwise and blockwise-32 recipes are known-bad (quality collapse / immediate EOS) and were not used.

Validation

9-clip Korean/English test set (Korean tactical-report sentence "2025년 3월 5일 전술평가 결과 보고", English equivalent, weather/status sentences, short Korean voice commands). Greedy decode, desktop CPU (XNNPACK, 8 threads). CER against punctuation-normalized references (space-removed); "exact" = normalized exact match.

i4 was the best tier in the producing project's entire ASR matrix (tiny/base/medium/large-v3/turbo × f32/i8/i4): 8/9 exact, character-perfect in both languages, and the fastest large-family tier per decode step — at 69 % of i8's size it beats i8 on every measured axis. i8's single character error is one short voice-command take heard as 볼륨 억 instead of 볼륨 업 (CER ko 0.000 on the other five Korean clips); its remaining misses vs exact are spacing-only.

i4 is additionally device-validated on Android arm64 (Snapdragon 865-class, shape-based decode tensor binding).

Runtime compatibility

Validated with the LiteRT (ai-edge-litert) interpreter on Windows x86_64 CPU (XNNPACK) and within the LiteRT-LM-Unity v0.14.0 pipeline (LiteRT-LM v0.14.0, Windows x86_64 + Android arm64, Snapdragon 865-class device).

Caveats

• 30 s fixed window; decoder is a full-sequence re-run per step (no KV cache) — a KV-cached runtime will be substantially faster per token.
• Requires the 128-mel frontend and 51866-vocab token ids (see integration notes).
• On Android the decode signature was bound by tensor shape rather than name in the producing project (tensor naming can differ across LiteRT versions) — bind decode inputs by shape if name-based binding fails.
• Language forcing recommended for short clips (e.g. <|ko|> / <|en|>).

Produced by

LiteRT-LM-Unity (release v0.14.0-unity). Whisper weights are MIT (OpenAI); quantized derivatives inherit the base license.