* Adding temperature scaling on Joiner logits:
- T hard-coded to 2.0
- so far best result NCE 0.122 (still not so high)
- the BPE scores were rescaled with 0.2 (but then also incorrect words
get high confidence, visually reasonable histograms are for 0.5 scale)
- BPE->WORD score merging done by min(.) function
(tried also prob-product, and also arithmetic, geometric, harmonic mean)
- without temperature scaling (i.e. scale 1.0), the best NCE was 0.032 (here product merging was best)
Results seem consistent with: https://arxiv.org/abs/2110.15222
Everything tuned on a very-small set of 100 sentences with 813 words and 10.2% WER, a Czech model.
I also experimented with blank posteriors mixed into the BPE confidences,
but no NCE improvement found, so not pushing that.
Temperature scling added also to the Greedy search confidences.
* making `temperature_scale` configurable from outside
207 lines
6.1 KiB
C++
207 lines
6.1 KiB
C++
// sherpa-onnx/csrc/online-recognizer.h
|
|
//
|
|
// Copyright (c) 2023 Xiaomi Corporation
|
|
|
|
#ifndef SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_H_
|
|
#define SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_H_
|
|
|
|
#include <memory>
|
|
#include <string>
|
|
#include <vector>
|
|
|
|
#if __ANDROID_API__ >= 9
|
|
#include "android/asset_manager.h"
|
|
#include "android/asset_manager_jni.h"
|
|
#endif
|
|
|
|
#include "sherpa-onnx/csrc/endpoint.h"
|
|
#include "sherpa-onnx/csrc/features.h"
|
|
#include "sherpa-onnx/csrc/online-ctc-fst-decoder-config.h"
|
|
#include "sherpa-onnx/csrc/online-lm-config.h"
|
|
#include "sherpa-onnx/csrc/online-model-config.h"
|
|
#include "sherpa-onnx/csrc/online-stream.h"
|
|
#include "sherpa-onnx/csrc/online-transducer-model-config.h"
|
|
#include "sherpa-onnx/csrc/parse-options.h"
|
|
|
|
namespace sherpa_onnx {
|
|
|
|
struct OnlineRecognizerResult {
|
|
/// Recognition results.
|
|
/// For English, it consists of space separated words.
|
|
/// For Chinese, it consists of Chinese words without spaces.
|
|
/// Example 1: "hello world"
|
|
/// Example 2: "你好世界"
|
|
std::string text;
|
|
|
|
/// Decoded results at the token level.
|
|
/// For instance, for BPE-based models it consists of a list of BPE tokens.
|
|
std::vector<std::string> tokens;
|
|
|
|
/// timestamps.size() == tokens.size()
|
|
/// timestamps[i] records the time in seconds when tokens[i] is decoded.
|
|
std::vector<float> timestamps;
|
|
|
|
std::vector<float> ys_probs; //< log-prob scores from ASR model
|
|
std::vector<float> lm_probs; //< log-prob scores from language model
|
|
//
|
|
/// log-domain scores from "hot-phrase" contextual boosting
|
|
std::vector<float> context_scores;
|
|
|
|
/// ID of this segment
|
|
/// When an endpoint is detected, it is incremented
|
|
int32_t segment = 0;
|
|
|
|
/// Starting time of this segment.
|
|
/// When an endpoint is detected, it will change
|
|
float start_time = 0;
|
|
|
|
/// True if the end of this segment is reached
|
|
bool is_final = false;
|
|
|
|
/** Return a json string.
|
|
*
|
|
* The returned string contains:
|
|
* {
|
|
* "text": "The recognition result",
|
|
* "tokens": [x, x, x],
|
|
* "timestamps": [x, x, x],
|
|
* "ys_probs": [x, x, x],
|
|
* "lm_probs": [x, x, x],
|
|
* "context_scores": [x, x, x],
|
|
* "segment": x,
|
|
* "start_time": x,
|
|
* "is_final": true|false
|
|
* }
|
|
*/
|
|
std::string AsJsonString() const;
|
|
};
|
|
|
|
struct OnlineRecognizerConfig {
|
|
FeatureExtractorConfig feat_config;
|
|
OnlineModelConfig model_config;
|
|
OnlineLMConfig lm_config;
|
|
EndpointConfig endpoint_config;
|
|
OnlineCtcFstDecoderConfig ctc_fst_decoder_config;
|
|
bool enable_endpoint = true;
|
|
|
|
std::string decoding_method = "greedy_search";
|
|
// now support modified_beam_search and greedy_search
|
|
|
|
// used only for modified_beam_search
|
|
int32_t max_active_paths = 4;
|
|
|
|
/// used only for modified_beam_search
|
|
float hotwords_score = 1.5;
|
|
std::string hotwords_file;
|
|
|
|
float blank_penalty = 0.0;
|
|
|
|
float temperature_scale = 2.0;
|
|
|
|
OnlineRecognizerConfig() = default;
|
|
|
|
OnlineRecognizerConfig(
|
|
const FeatureExtractorConfig &feat_config,
|
|
const OnlineModelConfig &model_config,
|
|
const OnlineLMConfig &lm_config,
|
|
const EndpointConfig &endpoint_config,
|
|
const OnlineCtcFstDecoderConfig &ctc_fst_decoder_config,
|
|
bool enable_endpoint,
|
|
const std::string &decoding_method,
|
|
int32_t max_active_paths,
|
|
const std::string &hotwords_file,
|
|
float hotwords_score,
|
|
float blank_penalty,
|
|
float temperature_scale)
|
|
: feat_config(feat_config),
|
|
model_config(model_config),
|
|
lm_config(lm_config),
|
|
endpoint_config(endpoint_config),
|
|
ctc_fst_decoder_config(ctc_fst_decoder_config),
|
|
enable_endpoint(enable_endpoint),
|
|
decoding_method(decoding_method),
|
|
max_active_paths(max_active_paths),
|
|
hotwords_file(hotwords_file),
|
|
hotwords_score(hotwords_score),
|
|
blank_penalty(blank_penalty),
|
|
temperature_scale(temperature_scale) {}
|
|
|
|
void Register(ParseOptions *po);
|
|
bool Validate() const;
|
|
|
|
std::string ToString() const;
|
|
};
|
|
|
|
class OnlineRecognizerImpl;
|
|
|
|
class OnlineRecognizer {
|
|
public:
|
|
explicit OnlineRecognizer(const OnlineRecognizerConfig &config);
|
|
|
|
#if __ANDROID_API__ >= 9
|
|
OnlineRecognizer(AAssetManager *mgr, const OnlineRecognizerConfig &config);
|
|
#endif
|
|
|
|
~OnlineRecognizer();
|
|
|
|
/// Create a stream for decoding.
|
|
std::unique_ptr<OnlineStream> CreateStream() const;
|
|
|
|
/** Create a stream for decoding.
|
|
*
|
|
* @param The hotwords for this string, it might contain several hotwords,
|
|
* the hotwords are separated by "/". In each of the hotwords, there
|
|
* are cjkchars or bpes, the bpe/cjkchar are separated by space (" ").
|
|
* For example, hotwords I LOVE YOU and HELLO WORLD, looks like:
|
|
*
|
|
* "▁I ▁LOVE ▁YOU/▁HE LL O ▁WORLD"
|
|
*/
|
|
std::unique_ptr<OnlineStream> CreateStream(const std::string &hotwords) const;
|
|
|
|
/**
|
|
* Return true if the given stream has enough frames for decoding.
|
|
* Return false otherwise
|
|
*/
|
|
bool IsReady(OnlineStream *s) const;
|
|
|
|
/** Decode a single stream. */
|
|
void DecodeStream(OnlineStream *s) const {
|
|
OnlineStream *ss[1] = {s};
|
|
DecodeStreams(ss, 1);
|
|
}
|
|
|
|
/**
|
|
* Warmups up onnxruntime sessions by apply optimization and
|
|
* allocating memory prior
|
|
*
|
|
* @param warmup Number of warmups.
|
|
* @param mbs : max-batch-size Max batch size for the models
|
|
*/
|
|
void WarmpUpRecognizer(int32_t warmup, int32_t mbs) const;
|
|
|
|
/** Decode multiple streams in parallel
|
|
*
|
|
* @param ss Pointer array containing streams to be decoded.
|
|
* @param n Number of streams in `ss`.
|
|
*/
|
|
void DecodeStreams(OnlineStream **ss, int32_t n) const;
|
|
|
|
OnlineRecognizerResult GetResult(OnlineStream *s) const;
|
|
|
|
// Return true if we detect an endpoint for this stream.
|
|
// Note: If this function returns true, you usually want to
|
|
// invoke Reset(s).
|
|
bool IsEndpoint(OnlineStream *s) const;
|
|
|
|
// Clear the state of this stream. If IsEndpoint(s) returns true,
|
|
// after calling this function, IsEndpoint(s) will return false
|
|
void Reset(OnlineStream *s) const;
|
|
|
|
private:
|
|
std::unique_ptr<OnlineRecognizerImpl> impl_;
|
|
};
|
|
|
|
} // namespace sherpa_onnx
|
|
|
|
#endif // SHERPA_ONNX_CSRC_ONLINE_RECOGNIZER_H_
|