seda-v0.2-tiny: streaming Turkish speech recognition in 4.5M parameters

seda-v0.2-tiny is a 15× smaller version of seda-v0.2 for slow CPUs, phones and the browser. It is a streaming Zipformer2 transducer with 4.5M parameters (seda-v0.2: 66M), distilled from seda-v0.2 and exported to int8 ONNX for sherpa-onnx. The acoustic model is about 7 MB.

• About as accurate as Whisper small on Turkish, with 50× fewer parameters and streaming output: 11.24 WER on FLEURS-TR test against 11.87 for Whisper small, and 16.88 against 23.25 on broadcast audio (MediaSpeech).
• Better than seda-v0.1 (66M parameters, 11.60 WER), while 15× smaller.
• Runs in a phone's browser: with sherpa-onnx WASM on one thread of a low-end Android phone, the 0.64 s chunk model with beam search runs at 0.77× real time.
• The same language model, voice activity detector and text formatter setup as seda-v0.2, so every sedalib mode works.

Usage

With sedalib 1.1 or newer:

pip install "sedalib[format]"  # without [format]: no punctuation, casing or digits

from sedalib import Recognizer, read_wav

samples, sample_rate = read_wav("audio.wav")

recognizer = Recognizer.from_pretrained("atasoglu/seda-v0.2-tiny", mode="fast")
print(recognizer.transcribe(samples, sample_rate))

For live audio, feed the stream small pieces as they arrive:

stream = recognizer.stream()

for chunk in microphone_chunks():  # e.g. 100 ms of 16 kHz audio
    print(stream.accept(chunk, sample_rate=16000))  # partial text
    if stream.is_endpoint:
        print("final:", stream.end_utterance())

from_pretrained downloads only the files of the chosen options: fast needs 7 MB, the language model adds 20 MB. The files are plain sherpa-onnx streaming transducer files, so sherpa_onnx.OnlineRecognizer.from_transducer loads them directly. The LM settings are modified_beam_search, lm_scale=0.6 and lodr_scale=-0.5.

Modes and chunk sizes

Mode	Decoding	FLEURS test WER
fast	0.64 s chunks, greedy	17.30
balanced (default)	1.28 s chunks, beam 4 + LM	11.24
accurate	0.64 s chunks live, then a second pass with full context, beam + LM	10.51

Shorter chunks react sooner but cost more CPU per second of audio and are a little less accurate. Pick them with chunk=:

Chunk	Directory	FLEURS test WER, greedy	with beam 4 + LM
0.16 s (chunk=8)	c8/	21.08	13.99
0.32 s (chunk=16)	c16/	19.20	12.85
0.64 s (chunk=32)	c32/	17.30	11.67
1.28 s (chunk=64)	root	16.50	11.24

• Measurement: sherpa-onnx int8 on CPU, audio followed by 1 s of silence. The accurate row decodes each utterance with full context (a 20.48 s chunk) and beam 4 + LM. The 0.16 and 0.32 s rows were measured with 0.64 s of left context; the released c8/ and c16/ use 1.28 s, which was 0.5 to 0.8 WER better on FLEURS dev.
• Slow devices: on a low-end Android phone (sherpa-onnx WASM, one thread), 0.16 s chunks with beam search needed 1.07 s of compute per second of audio and fell behind; 0.32 s greedy needed 0.89 s and 0.64 s with beam search 0.77 s. Use 0.32 or 0.64 s chunks there; 0.16 s suits fast CPUs where the earliest reaction matters.

Evaluation

WER (%). Lower is better. References and hypotheses were normalized the same way: lowercase, no punctuation, numbers spelled out. Whisper ran with faster-whisper on a GPU and decodes whole recordings offline; seda-v0.1 is in its released setup (0.64 s chunks, beam 4 + LM).

Test set	Whisper small (244M)	seda-v0.1 (66M)	seda-v0.2-tiny (4.5M) balanced	accurate	seda-v0.2 (66M) accurate
FLEURS-TR test (743 utt.)	11.87	11.60	11.24	10.51	6.38
MediaSpeech TR (broadcast, 10 h)	23.25	17.50	16.88 ¹		11.27

¹ 0.64 s chunks, beam 4 + LM.

More test sets, 0.64 s chunks:

Test set	Greedy	Beam 4 + LM	seda-v0.2 teacher, greedy
Khan Academy TR test ²	14.61	11.63	8.72
ISSAI TSC test	23.33	20.68	14.86

² Khan Academy is not in this model's training data (it is in seda-v0.2's).

The gap to seda-v0.2 is the price of 15× fewer parameters. The tiny model makes about 1.6× as many errors as its teacher. Use seda-v0.2 where it fits; use this model where it does not.

Model details

Architecture	Zipformer2 encoder + stateless decoder + joiner (pruned RNN-T), causal, trained with dynamic chunks
Parameters	4.49M
Encoder	6 stacks, layers 1,1,2,2,1,1, dim 128, feedforward 256, 4 heads
Input	80-dim log-mel fbank, 16 kHz mono (other rates are resampled)
Output	500 BPE tokens (the same as seda-v0.2). Lowercase spoken-form Turkish, without punctuation; numbers are words
Quantization	int8 (dynamic) encoder and joiner, fp32 decoder
Language model	seda-v0.2's LSTM 2×1024 RNN LM (int8, 17 MB) for shallow fusion, token bigram for LODR
Formatter	2-layer Turkish BERT (ytu-ce-cosmos/turkish-tiny-bert-uncased) with heads for punctuation, casing and apostrophes; int8, 4.7 MB

Files

File	Description
encoder.int8.onnx, decoder.onnx, joiner.int8.onnx, tokens.txt	Streaming model, 1.28 s chunks (balanced)
c32/, c16/, c8/	The same weights, exported with 0.64, 0.32 and 0.16 s chunks
final/	The same weights, exported with 5.12 s chunks and 20.48 s left context (second pass of accurate)
lm/rnnlm.int8.onnx, lm/2gram.fst	LSTM language model and token bigram for LODR, from seda-v0.2
vad/silero_vad.onnx	Silero VAD v6, 16 kHz (MIT, see vad/LICENSE)
formatter/	Text formatter (int8 ONNX token classifier) and its tokenizer

Training

• Knowledge distillation from seda-v0.2. The student learns the transcripts (pruned RNN-T loss) and, with weight 0.5, the teacher's encoder outputs (hidden-state distillation through a linear projection). The teacher's outputs were computed once for all training audio.
• Acoustic data: 1,913 h:
  • ISSAI Turkish Speech Corpus (209 h)
  • Alania synthetic speech, CC BY 4.0 subset (166 h)
  • Common Voice 17 (11 h)
  • YODAS2 Turkish, 1,527 h: YouTube videos under CC BY 3.0, labeled by Whisper large-v3-turbo and filtered by agreement with the video's subtitles. The labels are published as atasoglu/yodas-tr-whisper-labels.
  • Evaluation sentences were held out the same way as for seda-v0.2.
• Recipe:
  • Trained from scratch for 11 epochs: base LR 0.035, Eden schedule, fp16, max duration 1500 s, utterances up to 30 s.
  • Chunk sizes 0.16/0.32/0.64/1.28 s and full context, sampled per batch, so one set of weights serves every export.
  • MUSAN noise augmentation on the student's input only.
  • The release weights average the model over epochs 10 and 11.
• Language model and formatter: the LM is seda-v0.2's. The formatter is a 2-layer BERT fine-tuned on Wikipedia, FineWeb-2 and the YODAS Whisper labels, like seda-v0.2's 4-layer one.

Limitations

• Less accurate than seda-v0.2: about 1.6× the errors. Noisy, reverberant and telephone audio were not measured for this model.
• Foreign names and abbreviations are the main error source. The model often spells them by sound.
• The formatter places commas poorly (F1 0.38 on FLEURS dev); sentence ends, casing and apostrophes are close to seda-v0.2's.
• Turkish only, 16 kHz.

License

The model files are released under CC BY-SA 4.0:

• The language model and the formatter were trained on Wikipedia (CC BY-SA).
• The acoustic model's own training data has no share-alike source (Khan Academy, CC BY-SA 3.0, was removed), but it was distilled from seda-v0.2, whose training data includes Khan Academy.
• vad/silero_vad.onnx is Silero VAD, MIT license (vad/LICENSE).
• Other sources and their licenses: YODAS2 CC BY 3.0 (derived from YouTube videos by their creators); ISSAI Turkish Speech Corpus CC BY 4.0; Alania speech CC BY 4.0; Common Voice CC0; FineWeb-2 ODC-By 1.0; MUSAN CC BY 4.0 (augmentation only); Whisper MIT (labeler only); ytu-ce-cosmos/turkish-tiny-bert-uncased (YTU COSMOS) MIT, base of the formatter.

Acknowledgements and citation

Built with k2, icefall, lhotse, sherpa-onnx, Silero VAD and Whisper. The formatter builds on the Turkish BERT models of YTU COSMOS.

See the seda-v0.2 card for the dataset citations (YODAS, ISSAI TSC, Common Voice, MUSAN, Zipformer).